'
\nAnd now, the moment you've all been waiting for… the list of this year's most groundbreaking, earth-shattering, awe-inspiring academic research papers about GPS + Deep Learning!
\nDrumroll please…
\nEnjoy! <3
\nIdentification of travelers’ transportation modes is a fundamental step for various problems that arise in the domain of transportation such as travel demand analysis, transport planning, and traffic management. In this paper, we aim to identify travelers’ transportation modes purely based on their GPS trajectories. First, a segmentation process is developed to partition a user's trip into GPS segments with only one transportation mode. A majority of studies have proposed mode inference models based on hand-crafted features, which might be vulnerable to traffic and environmental conditions. Furthermore, the classification task in almost all models have been performed in a supervised fashion while a large amount of unlabeled GPS trajectories has remained unused. Accordingly, we propose a deep SEmi-Supervised Convolutional Autoencoder (SECA) architecture that can not only automatically extract relevant features from GPS segments but also exploit useful information in unlabeled data. The SECA integrates a convolutional-deconvolutional autoencoder and a convolutional neural network into a unified framework to concurrently perform supervised and unsupervised learning. The two components are simultaneously trained using both labeled and unlabeled GPS segments, which have already been converted into an efficient representation for the convolutional operation. An optimum schedule for varying the balancing parameters between reconstruction and classification errors are also implemented. The performance of the proposed SECA model, trip segmentation, the method for converting a raw trajectory into a new representation, the hyperparameter schedule, and the model configuration are evaluated by comparing to several baselines and alternatives for various amounts of labeled and unlabeled data. Our experimental results demonstrate the superiority of the proposed model over the state-of-the-art semi-supervised and supervised methods with respect to metrics such as accuracy and F-measure.
\nVehicle Localization During GPS Outages With Extended Kalman Filter and Deep Learning
\n\n\nTL;DR: use IEKF and a deep learning framework with multiple long short-term memory (multi-LSTM) modules
\n
Integration of microelectromechanical system-based inertial navigation system (MEMS-INS) and global positioning system (GPS) is a promising approach to vehicle localization. However, such a scheme may have poor performance during GPS outages and is less robust to measurement noises in changeable urban environments. In this article, we give an improved extended Kalman filter (IEKF) using an adaptation mechanism to eliminate the influence of noises in MEMS-INS and mitigate dependence on the process model. Especially, to guarantee accurate position estimation of the INS, a deep learning framework with multiple long short-term memory (multi-LSTM) modules is proposed to predict the increment of the vehicle position based on Gaussian mixture model (GMM) and Kullback–Leibler (KL) distance. The IEKF and the multi-LSTM are then combined together to optimize vehicle positioning accuracy during GPS outages in changeable urban environments. Numerical simulations and real-world experiments have demonstrated the effectiveness of the combined IEKF and multi-LSTM method, with the root-mean-square error (RMSE) of predicted position reduced by up to 93.9%. Or specifically, the RMSEs during GPS outages with durations 30, 60, and 120 s are 2.34, 2.69, and 3.08 m, respectively, which obviously outperform the existing method.
\nTravel Mode Identification With GPS Trajectories Using Wavelet Transform and Deep Learning
\nAccurate identification in public travel modes is an essential task in intelligent transportation systems. In recent years, GPS-based identification is gradually replacing the conventional survey-based information-gathering process due to the more detailed and precise data on individual’s travel patterns. Nonetheless, existing research suffers from deficient feature selection, high data dimensionality, and data under-utilization issues. In this work, we propose a novel travel mode identification mechanism based on discrete wavelet transform and recent developments of deep learning techniques. The proposed mechanism aims to take GPS trajectories of arbitrary lengths to develop accurate travel mode results in both global and online identification scenarios. In this mechanism, raw GPS data is first pre-processed to compute preliminary motion and displacement attributes, which are input into a tailor-made deep neural network. Discrete wavelet transform is also adopted to further extract time-frequency domain characteristics of the trajectories to assist the neural network in the classification task. To evaluate the performance of the proposed mechanism, a series of comprehensive case studies are conducted. The results indicate that the mechanism can notably outperform existing travel mode identifications on a same data set with minuscule computation time. Furthermore, an architecture test is performed to determine the best-performing structure for the proposed mechanism. Lastly, we demonstrate the capability of the mechanism in handling online identifications, and the performance sensitivity of the selected attributes is evaluated.
\n\n\nTL;DR: The deep learning models implemented using the vertical TEC (VTEC) time-series data estimated from GPS measurements over Bengaluru, Guntur, and Lucknow GPS stations. The LSTM-CNN model performs well when compared to other ionospheric deep learning forecasting algorithms with minimum root-mean-square error (RMSE) of 1.5 TEC units (TECUs) and a high degree of $R^{2} = 0.99$ .
\n
Prominent advances in the field of artificial intelligence during the past decade and the breakthrough of deep learning would be useful for investigating ionospheric weather using ground and space-based ionospheric sensors data. The significance of deep learning algorithms needs to be assessed in forecasting the low latitude ionospheric disturbances (delays) for the global positioning system (GPS) signals. Total electron content (TEC) data sets prepared by taking advantage of GPS satellite radio frequency (RF) signals. This letter provides the application of deep learning models, long short-term memory (LSTM), gated recurrent unit (GRU), and a hybrid model that consists of LSTM combined with convolution neural network (CNN) to forecast the ionospheric delays for GPS signals. The deep learning models implemented using the vertical TEC (VTEC) time-series data estimated from GPS measurements over Bengaluru, Guntur, and Lucknow GPS stations. The LSTM-CNN model performs well when compared to other ionospheric deep learning forecasting algorithms with minimum root-mean-square error (RMSE) of 1.5 TEC units (TECUs) and a high degree of $R^{2} = 0.99$ .
\nCoseismic displacement of Ahar–Varzegan earthquakes based on GPS observations and deep learning
\nThe determination of crustal deformation can be measured by geodetic observations of permanent global positioning system (GPS) stations. In this study, the coseismic displacement of 11 August 2012 with magnitudes 6.5 Mw and 6.3 Mw of Ahar–Varzegan earthquakes has been investigated based on GPS observations and deep learning. For this purpose, data were processed at a 30-s rate of 13 Iran geodynamic stations with distances of 25 to 160 km from the earthquake epicenter and then were entered into deep learning. The results show that the horizontal displacement field of the Ahar–Varzegan earthquake has a mean value of 27.93 cm and 15.35 cm, which is estimated with the root mean square error (RMSE) of ±0.24 cm. Vertical displacement has been neglected due to the low accuracy of the z component and the low density of stations in the central seismic range. Also, the right lateral fault (cause of Ahar–Varzegan earthquake) to seismic displacement is evident; field observations and previous research confirm coseismic displacement values and right latera fault.
\n\n\nTL;DR: We use deep learning to track objects in a game, and then we use GPS to keep track of the object's ID even when the objects overlap. This allows us to accurately track objects even in the presence of occlusion.
\n
When tracking objects (players, referees, etc.) in a game using deep learning, tracking often fails due to occlusion between objects. In this paper, we track the location of objects in the stadium through video tracking using deep learning. And we fused the GPS(Global Positioning System), which has a large error but can maintain the ID of the object even when the objects overlap so that the tracking can be done correctly even in the overlapping phenomenon between objects. From the experiment results, we could confirm that the object tracking failure rate can be reduced and the accuracy of the object location can be increased through the convergence of deep learning and GPS.
\nDeep Learning for GPS Spoofing Detection in Cellular-Enabled UAV Systems
\n\n\nTL;DR: Deep Learning of path loss measurements between aUAV and nearby base stations to detect spoofed GPS positions.
\n
Cellular-based Unmanned Aerial Vehicle (UAV) systems are a promising paradigm to provide reliable and fast Beyond Visual Line of Sight (BVLoS) communication services for UAV operations. However, such systems are facing a serious GPS spoofing threat for UAV’s position. To enable safe and secure UAV navigation BVLoS, this paper proposes a cellular network assisted UAV position monitoring and anti-GPS spoofing system, where deep learning approach is used to live detect spoofed GPS positions. Specifically, the proposed system introduces a MultiLayer Perceptron (MLP) model which is trained on the statistical properties of path loss measurements collected from nearby base stations to decide the authenticity of the GPS position. Experiment results indicate the accuracy rate of detecting GPS spoofing under our proposed approach is more than 93% with three base stations and it can also reach 80% with only one base station.
\nCapturing Uncertainty in Unsupervised GPS Trajectory Segmentation Using Bayesian Deep Learning
\nIntelligent transportation management requires not only statistical information on users’ mobility patterns, but also knowledge of their corresponding transportation modes. While GPS trajectories can be readily obtained from GPS sensors found in modern smartphones and vehicles, these massive geospatial data are neither automatically annotated nor segmented by transportation mode, subsequently complicating transportation mode identification. In addition, predictive uncertainty caused by the learned model parameters or variable noise in GPS sensor readings typically remains unaccounted for. To jointly address the above issues, we propose a Bayesian deep learning framework for unsupervised GPS trajectory segmentation. After unlabeled GPS trajectories are preprocessed into sequences of motion features, they are used in unsupervised training of a channel-calibrated temporal convolutional neural network for timestep-level transportation mode identification. At test time, we approximate variational inference via Monte Carlo dropout sampling, leveraging the mean and variance of the predicted distributions to classify each input timestep and estimate its predictive uncertainty, respectively. The proposed approach outperforms both its non-Bayesian variant and established GPS trajectory segmentation baselines on Microsoft’s Geolife dataset without using any labels.
\nA Novel Deep Learning GPS Anti-spoofing System with DOA Time-series Estimation
\n\n\nTL;DR: This paper reports a deep learning (DL)-based Global Positioning System (GPS) anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite. It uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals.
\n
Many critical systems and infrastructure rely on the Global Positioning System (GPS) for synchronization of clocks that is essential for their operation. Sophisticated GPS spoofing techniques can mimic legitimate GPS transmissions so closely making it difficult for even the most advanced anti-spoofing methods to detect them. This paper reports a novel deep learning (DL)-based GPS anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite. It uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals. A software implementation of the designed system using actual GPS orbital data demonstrated on average 95% accuracy even against dynamic airborne spoofing systems. A hardware implementation, using a 4-element antenna array, an RF-transceiver and a microprocessor, was shown to detect spoofed signals with above 93% accuracy. Unlike existing methods, this DL-based anti-spoofing system does not require knowledge of receiver's location and orientation or manual thresholds making it suitable for moving platforms. The proposed technique is capable of countering even the most advanced spoofing systems since it is difficult to exactly replicate the DOA time-series of a satellite even by an airborne spoofing transmitter.
\nA Novel Deep Learning GPS Anti-spoofing System with DOA Time-series Estimation
\n\n\nTL;DR: A deep learning-based GPS anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite is proposed. The technique uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals. The proposed technique is capable of countering even the most advanced spoofing systems
\n
Many critical systems and infrastructure rely on the Global Positioning System (GPS) for synchronization of clocks that is essential for their operation. Sophisticated GPS spoofing techniques can mimic legitimate GPS transmissions so closely making it difficult for even the most advanced anti-spoofing methods to detect them. This paper reports a novel deep learning (DL)-based GPS anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite. It uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals. A software implementation of the designed system using actual GPS orbital data demonstrated on average 95% accuracy even against dynamic airborne spoofing systems. A hardware implementation, using a 4-element antenna array, an RF-transceiver and a microprocessor, was shown to detect spoofed signals with above 93% accuracy. Unlike existing methods, this DL-based anti-spoofing system does not require knowledge of receiver's location and orientation or manual thresholds making it suitable for moving platforms. The proposed technique is capable of countering even the most advanced spoofing systems since it is difficult to exactly replicate the DOA time-series of a satellite even by an airborne spoofing transmitter.
\n\n\nTL;DR: use deep learning for video object tracking, then use GPS to improve accuracy for tracking overlapping objects
\n
When tracking objects (players, referees, etc.) in a game using deep learning, tracking often fails due to occlusion between objects. In this paper, we track the location of objects in the stadium through video tracking using deep learning. And we fused the GPS(Global Positioning System), which has a large error but can maintain the ID of the object even when the objects overlap so that the tracking can be done correctly even in the overlapping phenomenon between objects. From the experiment results, we could confirm that the object tracking failure rate can be reduced and the accuracy of the object location can be increased through the convergence of deep learning and GPS.
\nCapturing Uncertainty in Unsupervised GPS Trajectory Segmentation Using Bayesian Deep Learning
\n\n\n*TL;DR: *Bayesian deep learning framework for unsupervised GPS trajectory segmentation. Using mean and variance of the predicted distributions to classify each input timestep and estimate its predictive uncertainty, respectively.
\n
Intelligent transportation management requires not only statistical information on users’ mobility patterns, but also knowledge of their corresponding transportation modes. While GPS trajectories can be readily obtained from GPS sensors found in modern smartphones and vehicles, these massive geospatial data are neither automatically annotated nor segmented by transportation mode, subsequently complicating transportation mode identification. In addition, predictive uncertainty caused by the learned model parameters or variable noise in GPS sensor readings typically remains unaccounted for. To jointly address the above issues, we propose a Bayesian deep learning framework for unsupervised GPS trajectory segmentation. After unlabeled GPS trajectories are preprocessed into sequences of motion features, they are used in unsupervised training of a channel-calibrated temporal convolutional neural network for timestep-level transportation mode identification. At test time, we approximate variational inference via Monte Carlo dropout sampling, leveraging the mean and variance of the predicted distributions to classify each input timestep and estimate its predictive uncertainty, respectively. The proposed approach outperforms both its non-Bayesian variant and established GPS trajectory segmentation baselines on Microsoft’s Geolife dataset without using any labels.
\n\n\nTL;DR: use ICA or PCA in conjunction with LSTM to model ZTD for improved accuracy
\n
Tropospheric delay is an important error source in global positioning systems (GPS), and the water vapor retrieved from the tropospheric delay is widely used in meteorological research such as climate analysis and weather forecasting. Most zenith tropospheric delay (ZTD) models are presently used as positioning corrections, and few models are used for the estimation of water vapor, especially in Antarctica. Through two blind source separation algorithms (principal component analysis (PCA) and independent component analysis (ICA)), a back-propagation (BP) neural network and a deep learning technique (long short-term memory (LSTM) network), we establish an hourly high-accuracy ZTD model for GPS meteorology using the GPS-ZTD from 52 GPS stations in West Antarctica. Our results show that under the condition in which the principal components (PCs) and independent components (ICs) remain fixed after decomposition, the mean accuracy of the models for West Antarctica using PCA or ICA are better than 10 mm. Compared with the ZTDs from the nonmodeling stations, the mean root mean square (RMS) of the PCA and ICA models are 9.3 and 8.9 mm, respectively, and the correlation coefficients between the GPS-ZTD and model-ZTDs all exceed 90%. The accuracy of the ICA model is slightly higher than that of the PCA model, and the ICs of the ICA model show more consistent spatial responses. The six-hour forecast is the best among the forecast results, with a mean correlation coefficient of 90.6% and a mean RMS of 7.2 mm using GPS-ZTD. The long-term forecast result is significantly inaccurate, as the correlation coefficient between the 24-h forecast and GPS-ZTD is only 63.2%. Generally modest results have been achieved (HSS ≤ 0.38). Furthermore, the forecast accuracy in coastal areas is lower than that in inland areas. Our study confirms that the combined use of ICA and deep learning in ZTD modeling can effectively restore the original signals, and short-term forecasting can be effectively used in GPS meteorology. However, further development of the technology is necessary.
\nMultiMix: A Multi-Task Deep Learning Approach for Travel Mode Identification with Few GPS Data
\n\n\nTL;DR: we propose a semi-supervised framework for travel mode identification which outperforms existing methods, with accuracy reaching 66.2% on Geolife when just 1% of labeled data is used.
\n
Understanding how people choose to travel is essential for intelligent transportation planning and related smart services. Recent advances in deep learning, coupled with the increasing market penetration of GPS devices, have paved the way for novel travel mode identification methods based on GPS data mining. While many have shown promising results, most methods have often relied heavily on the few available labeled data, leaving large amounts of unlabeled ones unused. To address this issue, we propose MultiMix, a semi-supervised multi-task learning framework for travel mode identification. Our framework trains a deep autoencoder using batches of labeled, unlabeled, and synthetic data by simultaneously optimizing three corresponding objective functions. We show that MultiMix outperforms several fully-and semi-supervised baselines, achieving a classification accuracy of 66.2% on Geolife using just 1% of labeled data, with accuracy reaching 84.8% when incorporating all available labels. We also verify the necessity of its components through an ablation study designed to provide insights into the proposed approach.
\nEnhancement of GPS Position Accuracy Using Machine Vision and Deep Learning Techniques
\n\n\nTL;DR: Neural networks have been used to improve the accuracy of GPS position estimation in urban areas by using computer vision and deep learning methods. Empirical selection of hyper-parameters is necessary to obtain accurate results.
\n
The accuracy of GPS position estimation in urban cities is an issue which need to be resolved using machine vision and deep learning techniques. The accuracy of GPS in horizontal direction is better than in the vertical direction. Although for most of the navigation applications in intelligent transportation systems, horizontal positioning accuracy is vital, but vertical position accuracy gives idea about road slanting conditions. Several statistical methods like median filtering, homomorphic filtering and k-means clustering, etc., can be used to improve upon the position accuracy of GPS signals. Such methods are useful for offline applications where a lot many GPS measurements are taken at a single point and afterwards filtering is applied to batch of measurement. In this study, the GPS positioning errors which are caused by sensor noise, ionospheric effects, occlusions by building facades, etc., have been considered for online improvement in position estimation using computer vision and deep learning methods by empirically choosing hyper-parameters.
\nA Deep Learning-Based Approach to Forecast Ionospheric Delays for GPS Signals
\nThis letter proposes the implementation of ionospheric forecasting model based on the long short-term memory (LSTM) networks. Ionospheric region produces time delay for radio wave propagation of global positioning system (GPS) satellites. The ionospheric delays for GPS signals degrade the position accuracy in the measurements for precise navigation and positioning services. Utilizing the emerging artificial intelligence mathematical tools to forecast ionospheric disturbances using GPS-estimated total electron content (TEC) observations is decisive. In this letter, multi-input LSTM forecasting technique is investigated and tested for evaluating its capability in forecasting the ionospheric delays over Bengaluru station (16.26° N, 80.44° E) using eight years (2009–2016) of GPS measured vertical TEC (VTEC) time-series data. The assessment of the LSTM model performance during geomagnetic quiet and disturbed conditions is carried out in comparison with artificial neural networks model and International Reference Ionosphere (IRI-2016) model based on statistical parameters like root-mean-square error and coefficient of determination ( $R^{2}$ ). The experimental analysis delineates that the proposed LSTM model has provided the correlation of 0.99 with the GPS-measured VTEC and with a forecasting error of 1–2 TEC units.
\n\n\nTL;DR: A GPS map-matching algorithm is proposed to accurately match the data points generated by trucks driving on urban express roads. The data points are then screened and processed to remove abnormal values, and the resulting traffic speed sequence is used to train and test a GRU model. The accuracies of the proposed methodology are verified across four scenarios: workday, weekend, rainy, and accident conditions.
\n
Due to the restriction of traffic management measure in large cities, large heavy-haul trucks can only travel on the circuits and expressways around the city, which often causes congestion in these areas. It is necessary to study the travel speed prediction of trucks on the urban ring road and provide special information services for trucks. Based on the data generated by the trucks driving on the Sixth Ring Road in Beijing, an optimized GRU algorithm is proposed to predict the travel speed of trucks driving on urban express roads under non-recurrent congested conditions. First, a GPS map-matching algorithm that can simultaneously meet the accuracy and efficiency requirements of matching is proposed. Then, the trucks’ data traveling on the Sixth Ring Road in Beijing are extracted from the original data. Aiming at getting rid of the abnormal data in GPS data, the screening and processing rules of the abnormal data are made, and then, the traffic speed sequence is extracted. Aiming at the problem that the commonly used weight optimization algorithm SGD cannot adaptively adjust the learning rate, Adam, Adadelta, and Rmsprop are used to optimize the weights in the GRU model in this paper. Considering the four scenarios, including workday, weekend, rainy, and accident, the accuracies of the proposed methods are verified.
\nClassical machine learning and deep neural network ensemble model for GPS-based activity recognition
\nOur KDDI Research team proposes an ensemble model for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge using GPS data. During preprocessing, we corrected the GPS dataset that contains errors and missing values using a Kalman smoother. Since this smoother can be processed offline, it can be used to correct backforward using the recorded future locations in the dataset. Further, by using features that have similar distributions in the train subject’s data and the other subjects’ datasets, we could achieve robust feature selection across multiple subjects. The first stage of our ensemble model employs classical machine learning and deep neural network approaches independently, specifically, LightGBM and LSTM, respectively. The second stage calculates the weighted average of the outputs of both approaches. Our results show the improved accuracy contributed by our ensemble, suggesting that it effectively makes use of both statistical and non-statistical features given the suitable base models. We confirmed that the use of Kalman-smoother, selection of features with similar distributions across subjects and ensemble modeling contributed to improving the accuracy of both train and validation datasets.
\nDeep Learning Based Autonomous Vehicle Super Resolution DOA Estimation for Safety Driving
\n\n\nTL;DR: Autonomous vehicles can be located using the deep unfolding technique, which is a recent advance of deep learning. This approach can achieve higher accuracy than traditional methods such as GPS, and is also suitable for fast-moving vehicles.
\n
In this paper, a novel system architecture including a massive multi-input multi-output (MIMO) or a reconfigurable intelligent surface (RIS) and multiple autonomous vehicles is considered in vehicle location systems. The location parameters of autonomous vehicles can be estimated based on the deep unfolding technique, which is a recent advance of deep learning. Traditional vehicle location methods such as the global position system (GPS) can only locate the target vehicles with relatively low accuracy. The super resolution cannot be achieved when two vehicles are too close, which means that the safety incidents exist when autonomous vehicles are deployed in future intelligent transportation systems (ITS). Different from the existing massive MIMO or RIS equipped with a regular array such as uniform rectangular array (URA) and uniform circular array (UCA), we exploit a massive MIMO or a RIS equipped with a conformal array extended from traditional regular array. First, the rotation from the global coordinate system to the local coordinate system is achieved based on geometric algebra. Second, 2D-DOA estimation of autonomous vehicles is modeled as a novel block sparse recovery problem. Third, the deep network architecture SBLNet is implemented to learn the nonlinear characteristic from the DOAs of autonomous vehicles and the data received by massive MIMOs or RISs. The 2D-DOA and polarization parameters can be estimated based on SBLNet with relatively low computational complexity. Simulation results demonstrate that SBLNet performs better than the state-of-the-art methods in terms of estimation accuracy and successful probability. The SBLNet is also suitable for the practical scenario considering fast moving autonomous vehicles, while, the traditional block sparse recovery methods fail in this complex scenario.
\n'
\nAnd now, the moment you've all been waiting for… the list of this year's most groundbreaking, earth-shattering, awe-inspiring academic research papers about GPS + Deep Learning!
\nDrumroll please…
\nEnjoy! <3
\nIdentification of travelers’ transportation modes is a fundamental step for various problems that arise in the domain of transportation such as travel demand analysis, transport planning, and traffic management. In this paper, we aim to identify travelers’ transportation modes purely based on their GPS trajectories. First, a segmentation process is developed to partition a user's trip into GPS segments with only one transportation mode. A majority of studies have proposed mode inference models based on hand-crafted features, which might be vulnerable to traffic and environmental conditions. Furthermore, the classification task in almost all models have been performed in a supervised fashion while a large amount of unlabeled GPS trajectories has remained unused. Accordingly, we propose a deep SEmi-Supervised Convolutional Autoencoder (SECA) architecture that can not only automatically extract relevant features from GPS segments but also exploit useful information in unlabeled data. The SECA integrates a convolutional-deconvolutional autoencoder and a convolutional neural network into a unified framework to concurrently perform supervised and unsupervised learning. The two components are simultaneously trained using both labeled and unlabeled GPS segments, which have already been converted into an efficient representation for the convolutional operation. An optimum schedule for varying the balancing parameters between reconstruction and classification errors are also implemented. The performance of the proposed SECA model, trip segmentation, the method for converting a raw trajectory into a new representation, the hyperparameter schedule, and the model configuration are evaluated by comparing to several baselines and alternatives for various amounts of labeled and unlabeled data. Our experimental results demonstrate the superiority of the proposed model over the state-of-the-art semi-supervised and supervised methods with respect to metrics such as accuracy and F-measure.
\nVehicle Localization During GPS Outages With Extended Kalman Filter and Deep Learning
\n\n\nTL;DR: use IEKF and a deep learning framework with multiple long short-term memory (multi-LSTM) modules
\n
Integration of microelectromechanical system-based inertial navigation system (MEMS-INS) and global positioning system (GPS) is a promising approach to vehicle localization. However, such a scheme may have poor performance during GPS outages and is less robust to measurement noises in changeable urban environments. In this article, we give an improved extended Kalman filter (IEKF) using an adaptation mechanism to eliminate the influence of noises in MEMS-INS and mitigate dependence on the process model. Especially, to guarantee accurate position estimation of the INS, a deep learning framework with multiple long short-term memory (multi-LSTM) modules is proposed to predict the increment of the vehicle position based on Gaussian mixture model (GMM) and Kullback–Leibler (KL) distance. The IEKF and the multi-LSTM are then combined together to optimize vehicle positioning accuracy during GPS outages in changeable urban environments. Numerical simulations and real-world experiments have demonstrated the effectiveness of the combined IEKF and multi-LSTM method, with the root-mean-square error (RMSE) of predicted position reduced by up to 93.9%. Or specifically, the RMSEs during GPS outages with durations 30, 60, and 120 s are 2.34, 2.69, and 3.08 m, respectively, which obviously outperform the existing method.
\nTravel Mode Identification With GPS Trajectories Using Wavelet Transform and Deep Learning
\nAccurate identification in public travel modes is an essential task in intelligent transportation systems. In recent years, GPS-based identification is gradually replacing the conventional survey-based information-gathering process due to the more detailed and precise data on individual’s travel patterns. Nonetheless, existing research suffers from deficient feature selection, high data dimensionality, and data under-utilization issues. In this work, we propose a novel travel mode identification mechanism based on discrete wavelet transform and recent developments of deep learning techniques. The proposed mechanism aims to take GPS trajectories of arbitrary lengths to develop accurate travel mode results in both global and online identification scenarios. In this mechanism, raw GPS data is first pre-processed to compute preliminary motion and displacement attributes, which are input into a tailor-made deep neural network. Discrete wavelet transform is also adopted to further extract time-frequency domain characteristics of the trajectories to assist the neural network in the classification task. To evaluate the performance of the proposed mechanism, a series of comprehensive case studies are conducted. The results indicate that the mechanism can notably outperform existing travel mode identifications on a same data set with minuscule computation time. Furthermore, an architecture test is performed to determine the best-performing structure for the proposed mechanism. Lastly, we demonstrate the capability of the mechanism in handling online identifications, and the performance sensitivity of the selected attributes is evaluated.
\n\n\nTL;DR: The deep learning models implemented using the vertical TEC (VTEC) time-series data estimated from GPS measurements over Bengaluru, Guntur, and Lucknow GPS stations. The LSTM-CNN model performs well when compared to other ionospheric deep learning forecasting algorithms with minimum root-mean-square error (RMSE) of 1.5 TEC units (TECUs) and a high degree of $R^{2} = 0.99$ .
\n
Prominent advances in the field of artificial intelligence during the past decade and the breakthrough of deep learning would be useful for investigating ionospheric weather using ground and space-based ionospheric sensors data. The significance of deep learning algorithms needs to be assessed in forecasting the low latitude ionospheric disturbances (delays) for the global positioning system (GPS) signals. Total electron content (TEC) data sets prepared by taking advantage of GPS satellite radio frequency (RF) signals. This letter provides the application of deep learning models, long short-term memory (LSTM), gated recurrent unit (GRU), and a hybrid model that consists of LSTM combined with convolution neural network (CNN) to forecast the ionospheric delays for GPS signals. The deep learning models implemented using the vertical TEC (VTEC) time-series data estimated from GPS measurements over Bengaluru, Guntur, and Lucknow GPS stations. The LSTM-CNN model performs well when compared to other ionospheric deep learning forecasting algorithms with minimum root-mean-square error (RMSE) of 1.5 TEC units (TECUs) and a high degree of $R^{2} = 0.99$ .
\nCoseismic displacement of Ahar–Varzegan earthquakes based on GPS observations and deep learning
\nThe determination of crustal deformation can be measured by geodetic observations of permanent global positioning system (GPS) stations. In this study, the coseismic displacement of 11 August 2012 with magnitudes 6.5 Mw and 6.3 Mw of Ahar–Varzegan earthquakes has been investigated based on GPS observations and deep learning. For this purpose, data were processed at a 30-s rate of 13 Iran geodynamic stations with distances of 25 to 160 km from the earthquake epicenter and then were entered into deep learning. The results show that the horizontal displacement field of the Ahar–Varzegan earthquake has a mean value of 27.93 cm and 15.35 cm, which is estimated with the root mean square error (RMSE) of ±0.24 cm. Vertical displacement has been neglected due to the low accuracy of the z component and the low density of stations in the central seismic range. Also, the right lateral fault (cause of Ahar–Varzegan earthquake) to seismic displacement is evident; field observations and previous research confirm coseismic displacement values and right latera fault.
\n\n\nTL;DR: We use deep learning to track objects in a game, and then we use GPS to keep track of the object's ID even when the objects overlap. This allows us to accurately track objects even in the presence of occlusion.
\n
When tracking objects (players, referees, etc.) in a game using deep learning, tracking often fails due to occlusion between objects. In this paper, we track the location of objects in the stadium through video tracking using deep learning. And we fused the GPS(Global Positioning System), which has a large error but can maintain the ID of the object even when the objects overlap so that the tracking can be done correctly even in the overlapping phenomenon between objects. From the experiment results, we could confirm that the object tracking failure rate can be reduced and the accuracy of the object location can be increased through the convergence of deep learning and GPS.
\nDeep Learning for GPS Spoofing Detection in Cellular-Enabled UAV Systems
\n\n\nTL;DR: Deep Learning of path loss measurements between aUAV and nearby base stations to detect spoofed GPS positions.
\n
Cellular-based Unmanned Aerial Vehicle (UAV) systems are a promising paradigm to provide reliable and fast Beyond Visual Line of Sight (BVLoS) communication services for UAV operations. However, such systems are facing a serious GPS spoofing threat for UAV’s position. To enable safe and secure UAV navigation BVLoS, this paper proposes a cellular network assisted UAV position monitoring and anti-GPS spoofing system, where deep learning approach is used to live detect spoofed GPS positions. Specifically, the proposed system introduces a MultiLayer Perceptron (MLP) model which is trained on the statistical properties of path loss measurements collected from nearby base stations to decide the authenticity of the GPS position. Experiment results indicate the accuracy rate of detecting GPS spoofing under our proposed approach is more than 93% with three base stations and it can also reach 80% with only one base station.
\nCapturing Uncertainty in Unsupervised GPS Trajectory Segmentation Using Bayesian Deep Learning
\nIntelligent transportation management requires not only statistical information on users’ mobility patterns, but also knowledge of their corresponding transportation modes. While GPS trajectories can be readily obtained from GPS sensors found in modern smartphones and vehicles, these massive geospatial data are neither automatically annotated nor segmented by transportation mode, subsequently complicating transportation mode identification. In addition, predictive uncertainty caused by the learned model parameters or variable noise in GPS sensor readings typically remains unaccounted for. To jointly address the above issues, we propose a Bayesian deep learning framework for unsupervised GPS trajectory segmentation. After unlabeled GPS trajectories are preprocessed into sequences of motion features, they are used in unsupervised training of a channel-calibrated temporal convolutional neural network for timestep-level transportation mode identification. At test time, we approximate variational inference via Monte Carlo dropout sampling, leveraging the mean and variance of the predicted distributions to classify each input timestep and estimate its predictive uncertainty, respectively. The proposed approach outperforms both its non-Bayesian variant and established GPS trajectory segmentation baselines on Microsoft’s Geolife dataset without using any labels.
\nA Novel Deep Learning GPS Anti-spoofing System with DOA Time-series Estimation
\n\n\nTL;DR: This paper reports a deep learning (DL)-based Global Positioning System (GPS) anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite. It uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals.
\n
Many critical systems and infrastructure rely on the Global Positioning System (GPS) for synchronization of clocks that is essential for their operation. Sophisticated GPS spoofing techniques can mimic legitimate GPS transmissions so closely making it difficult for even the most advanced anti-spoofing methods to detect them. This paper reports a novel deep learning (DL)-based GPS anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite. It uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals. A software implementation of the designed system using actual GPS orbital data demonstrated on average 95% accuracy even against dynamic airborne spoofing systems. A hardware implementation, using a 4-element antenna array, an RF-transceiver and a microprocessor, was shown to detect spoofed signals with above 93% accuracy. Unlike existing methods, this DL-based anti-spoofing system does not require knowledge of receiver's location and orientation or manual thresholds making it suitable for moving platforms. The proposed technique is capable of countering even the most advanced spoofing systems since it is difficult to exactly replicate the DOA time-series of a satellite even by an airborne spoofing transmitter.
\nA Novel Deep Learning GPS Anti-spoofing System with DOA Time-series Estimation
\n\n\nTL;DR: A deep learning-based GPS anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite is proposed. The technique uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals. The proposed technique is capable of countering even the most advanced spoofing systems
\n
Many critical systems and infrastructure rely on the Global Positioning System (GPS) for synchronization of clocks that is essential for their operation. Sophisticated GPS spoofing techniques can mimic legitimate GPS transmissions so closely making it difficult for even the most advanced anti-spoofing methods to detect them. This paper reports a novel deep learning (DL)-based GPS anti-spoofing technique that relies only on the physical attributes unique to a signal originated at an orbiting satellite. It uses a multi-element antenna array receiver to estimate the instantaneous signal direction-of-arrival (DOA) using a subspace-based statistical signal processing. Time-series of the estimated DOAs of signals are input to a convolutional neural network deep-learning classifier that learns the embedded signatures unique to the trajectories of signal sources to separate authentic and spoofed GPS signals. A software implementation of the designed system using actual GPS orbital data demonstrated on average 95% accuracy even against dynamic airborne spoofing systems. A hardware implementation, using a 4-element antenna array, an RF-transceiver and a microprocessor, was shown to detect spoofed signals with above 93% accuracy. Unlike existing methods, this DL-based anti-spoofing system does not require knowledge of receiver's location and orientation or manual thresholds making it suitable for moving platforms. The proposed technique is capable of countering even the most advanced spoofing systems since it is difficult to exactly replicate the DOA time-series of a satellite even by an airborne spoofing transmitter.
\n\n\nTL;DR: use deep learning for video object tracking, then use GPS to improve accuracy for tracking overlapping objects
\n
When tracking objects (players, referees, etc.) in a game using deep learning, tracking often fails due to occlusion between objects. In this paper, we track the location of objects in the stadium through video tracking using deep learning. And we fused the GPS(Global Positioning System), which has a large error but can maintain the ID of the object even when the objects overlap so that the tracking can be done correctly even in the overlapping phenomenon between objects. From the experiment results, we could confirm that the object tracking failure rate can be reduced and the accuracy of the object location can be increased through the convergence of deep learning and GPS.
\nCapturing Uncertainty in Unsupervised GPS Trajectory Segmentation Using Bayesian Deep Learning
\n\n\n*TL;DR: *Bayesian deep learning framework for unsupervised GPS trajectory segmentation. Using mean and variance of the predicted distributions to classify each input timestep and estimate its predictive uncertainty, respectively.
\n
Intelligent transportation management requires not only statistical information on users’ mobility patterns, but also knowledge of their corresponding transportation modes. While GPS trajectories can be readily obtained from GPS sensors found in modern smartphones and vehicles, these massive geospatial data are neither automatically annotated nor segmented by transportation mode, subsequently complicating transportation mode identification. In addition, predictive uncertainty caused by the learned model parameters or variable noise in GPS sensor readings typically remains unaccounted for. To jointly address the above issues, we propose a Bayesian deep learning framework for unsupervised GPS trajectory segmentation. After unlabeled GPS trajectories are preprocessed into sequences of motion features, they are used in unsupervised training of a channel-calibrated temporal convolutional neural network for timestep-level transportation mode identification. At test time, we approximate variational inference via Monte Carlo dropout sampling, leveraging the mean and variance of the predicted distributions to classify each input timestep and estimate its predictive uncertainty, respectively. The proposed approach outperforms both its non-Bayesian variant and established GPS trajectory segmentation baselines on Microsoft’s Geolife dataset without using any labels.
\n\n\nTL;DR: use ICA or PCA in conjunction with LSTM to model ZTD for improved accuracy
\n
Tropospheric delay is an important error source in global positioning systems (GPS), and the water vapor retrieved from the tropospheric delay is widely used in meteorological research such as climate analysis and weather forecasting. Most zenith tropospheric delay (ZTD) models are presently used as positioning corrections, and few models are used for the estimation of water vapor, especially in Antarctica. Through two blind source separation algorithms (principal component analysis (PCA) and independent component analysis (ICA)), a back-propagation (BP) neural network and a deep learning technique (long short-term memory (LSTM) network), we establish an hourly high-accuracy ZTD model for GPS meteorology using the GPS-ZTD from 52 GPS stations in West Antarctica. Our results show that under the condition in which the principal components (PCs) and independent components (ICs) remain fixed after decomposition, the mean accuracy of the models for West Antarctica using PCA or ICA are better than 10 mm. Compared with the ZTDs from the nonmodeling stations, the mean root mean square (RMS) of the PCA and ICA models are 9.3 and 8.9 mm, respectively, and the correlation coefficients between the GPS-ZTD and model-ZTDs all exceed 90%. The accuracy of the ICA model is slightly higher than that of the PCA model, and the ICs of the ICA model show more consistent spatial responses. The six-hour forecast is the best among the forecast results, with a mean correlation coefficient of 90.6% and a mean RMS of 7.2 mm using GPS-ZTD. The long-term forecast result is significantly inaccurate, as the correlation coefficient between the 24-h forecast and GPS-ZTD is only 63.2%. Generally modest results have been achieved (HSS ≤ 0.38). Furthermore, the forecast accuracy in coastal areas is lower than that in inland areas. Our study confirms that the combined use of ICA and deep learning in ZTD modeling can effectively restore the original signals, and short-term forecasting can be effectively used in GPS meteorology. However, further development of the technology is necessary.
\nMultiMix: A Multi-Task Deep Learning Approach for Travel Mode Identification with Few GPS Data
\n\n\nTL;DR: we propose a semi-supervised framework for travel mode identification which outperforms existing methods, with accuracy reaching 66.2% on Geolife when just 1% of labeled data is used.
\n
Understanding how people choose to travel is essential for intelligent transportation planning and related smart services. Recent advances in deep learning, coupled with the increasing market penetration of GPS devices, have paved the way for novel travel mode identification methods based on GPS data mining. While many have shown promising results, most methods have often relied heavily on the few available labeled data, leaving large amounts of unlabeled ones unused. To address this issue, we propose MultiMix, a semi-supervised multi-task learning framework for travel mode identification. Our framework trains a deep autoencoder using batches of labeled, unlabeled, and synthetic data by simultaneously optimizing three corresponding objective functions. We show that MultiMix outperforms several fully-and semi-supervised baselines, achieving a classification accuracy of 66.2% on Geolife using just 1% of labeled data, with accuracy reaching 84.8% when incorporating all available labels. We also verify the necessity of its components through an ablation study designed to provide insights into the proposed approach.
\nEnhancement of GPS Position Accuracy Using Machine Vision and Deep Learning Techniques
\n\n\nTL;DR: Neural networks have been used to improve the accuracy of GPS position estimation in urban areas by using computer vision and deep learning methods. Empirical selection of hyper-parameters is necessary to obtain accurate results.
\n
The accuracy of GPS position estimation in urban cities is an issue which need to be resolved using machine vision and deep learning techniques. The accuracy of GPS in horizontal direction is better than in the vertical direction. Although for most of the navigation applications in intelligent transportation systems, horizontal positioning accuracy is vital, but vertical position accuracy gives idea about road slanting conditions. Several statistical methods like median filtering, homomorphic filtering and k-means clustering, etc., can be used to improve upon the position accuracy of GPS signals. Such methods are useful for offline applications where a lot many GPS measurements are taken at a single point and afterwards filtering is applied to batch of measurement. In this study, the GPS positioning errors which are caused by sensor noise, ionospheric effects, occlusions by building facades, etc., have been considered for online improvement in position estimation using computer vision and deep learning methods by empirically choosing hyper-parameters.
\nA Deep Learning-Based Approach to Forecast Ionospheric Delays for GPS Signals
\nThis letter proposes the implementation of ionospheric forecasting model based on the long short-term memory (LSTM) networks. Ionospheric region produces time delay for radio wave propagation of global positioning system (GPS) satellites. The ionospheric delays for GPS signals degrade the position accuracy in the measurements for precise navigation and positioning services. Utilizing the emerging artificial intelligence mathematical tools to forecast ionospheric disturbances using GPS-estimated total electron content (TEC) observations is decisive. In this letter, multi-input LSTM forecasting technique is investigated and tested for evaluating its capability in forecasting the ionospheric delays over Bengaluru station (16.26° N, 80.44° E) using eight years (2009–2016) of GPS measured vertical TEC (VTEC) time-series data. The assessment of the LSTM model performance during geomagnetic quiet and disturbed conditions is carried out in comparison with artificial neural networks model and International Reference Ionosphere (IRI-2016) model based on statistical parameters like root-mean-square error and coefficient of determination ( $R^{2}$ ). The experimental analysis delineates that the proposed LSTM model has provided the correlation of 0.99 with the GPS-measured VTEC and with a forecasting error of 1–2 TEC units.
\n\n\nTL;DR: A GPS map-matching algorithm is proposed to accurately match the data points generated by trucks driving on urban express roads. The data points are then screened and processed to remove abnormal values, and the resulting traffic speed sequence is used to train and test a GRU model. The accuracies of the proposed methodology are verified across four scenarios: workday, weekend, rainy, and accident conditions.
\n
Due to the restriction of traffic management measure in large cities, large heavy-haul trucks can only travel on the circuits and expressways around the city, which often causes congestion in these areas. It is necessary to study the travel speed prediction of trucks on the urban ring road and provide special information services for trucks. Based on the data generated by the trucks driving on the Sixth Ring Road in Beijing, an optimized GRU algorithm is proposed to predict the travel speed of trucks driving on urban express roads under non-recurrent congested conditions. First, a GPS map-matching algorithm that can simultaneously meet the accuracy and efficiency requirements of matching is proposed. Then, the trucks’ data traveling on the Sixth Ring Road in Beijing are extracted from the original data. Aiming at getting rid of the abnormal data in GPS data, the screening and processing rules of the abnormal data are made, and then, the traffic speed sequence is extracted. Aiming at the problem that the commonly used weight optimization algorithm SGD cannot adaptively adjust the learning rate, Adam, Adadelta, and Rmsprop are used to optimize the weights in the GRU model in this paper. Considering the four scenarios, including workday, weekend, rainy, and accident, the accuracies of the proposed methods are verified.
\nClassical machine learning and deep neural network ensemble model for GPS-based activity recognition
\nOur KDDI Research team proposes an ensemble model for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge using GPS data. During preprocessing, we corrected the GPS dataset that contains errors and missing values using a Kalman smoother. Since this smoother can be processed offline, it can be used to correct backforward using the recorded future locations in the dataset. Further, by using features that have similar distributions in the train subject’s data and the other subjects’ datasets, we could achieve robust feature selection across multiple subjects. The first stage of our ensemble model employs classical machine learning and deep neural network approaches independently, specifically, LightGBM and LSTM, respectively. The second stage calculates the weighted average of the outputs of both approaches. Our results show the improved accuracy contributed by our ensemble, suggesting that it effectively makes use of both statistical and non-statistical features given the suitable base models. We confirmed that the use of Kalman-smoother, selection of features with similar distributions across subjects and ensemble modeling contributed to improving the accuracy of both train and validation datasets.
\nDeep Learning Based Autonomous Vehicle Super Resolution DOA Estimation for Safety Driving
\n\n\nTL;DR: Autonomous vehicles can be located using the deep unfolding technique, which is a recent advance of deep learning. This approach can achieve higher accuracy than traditional methods such as GPS, and is also suitable for fast-moving vehicles.
\n
In this paper, a novel system architecture including a massive multi-input multi-output (MIMO) or a reconfigurable intelligent surface (RIS) and multiple autonomous vehicles is considered in vehicle location systems. The location parameters of autonomous vehicles can be estimated based on the deep unfolding technique, which is a recent advance of deep learning. Traditional vehicle location methods such as the global position system (GPS) can only locate the target vehicles with relatively low accuracy. The super resolution cannot be achieved when two vehicles are too close, which means that the safety incidents exist when autonomous vehicles are deployed in future intelligent transportation systems (ITS). Different from the existing massive MIMO or RIS equipped with a regular array such as uniform rectangular array (URA) and uniform circular array (UCA), we exploit a massive MIMO or a RIS equipped with a conformal array extended from traditional regular array. First, the rotation from the global coordinate system to the local coordinate system is achieved based on geometric algebra. Second, 2D-DOA estimation of autonomous vehicles is modeled as a novel block sparse recovery problem. Third, the deep network architecture SBLNet is implemented to learn the nonlinear characteristic from the DOAs of autonomous vehicles and the data received by massive MIMOs or RISs. The 2D-DOA and polarization parameters can be estimated based on SBLNet with relatively low computational complexity. Simulation results demonstrate that SBLNet performs better than the state-of-the-art methods in terms of estimation accuracy and successful probability. The SBLNet is also suitable for the practical scenario considering fast moving autonomous vehicles, while, the traditional block sparse recovery methods fail in this complex scenario.
\n