![\"Imgur\"](\"https://i.imgur.com/pAhVzTP.png\")
First of all, thank you very much for hosting this competition. Here we describe the team TOKYO's solution.
\nUnfortunately, our solution went through a shake down （Public 8th Private 19th）. We guess one of the reasons is that we targeted downtown & tree areas too much, while highway areas consist the major part of the private dataset.
\nOur solution consists of several algorithms, some of them introduced by other competitors and made public, the others developed by ourselves.
\n
The first module is a position estimation based on GNSS data provided in *_derived.csv. With some simple tricks as follow, we achived LB: 6.665 in public.
\nHere we trained a model which estimates the difference between estimated position \\((lat^i_{est}, lng^i_{est})\\) and corresponding ground truth position \\((lat^i_{gt}, lng^i_{gt})\\), based on the provided train data. We introduced several features and used LightGBM for the estimation.
\nBased on the idea introduced here by @dehokanta, we introduced a rule-based outlier detection algorithm. Through several experiments, we've found out that the estimation accuracy improved significantly by rejecting the following points:
\nWe used several types of phones mean method. Extended from the idea introduced here, we used several types of phone means such as:
\nAs discussed in several discussions & notebooks, the noise during the vehicle's stop was apparently one of the major noise in the baseline estimation.
\nWe developed stop detection algorithm using inertial measurements in a rule-based fashion.
For some collections which does not include sensor data, we estimated the stop state only from the position data.
\nWe used snap-to-grid for tree area. For downtown area, we applied a different snap-to-grid method described below.
\nTo overcome some severe problems in simple snap-to-grid, we combined the algorithm with Dynamic Time Warping （DTW）. This algorithm significantly worked especially for the downtown area.
\nWe used a method introduced here by the other team.
\nFirst of all, thank you very much for hosting this competition. Here we describe the team TOKYO's solution.
\nUnfortunately, our solution went through a shake down （Public 8th Private 19th）. We guess one of the reasons is that we targeted downtown & tree areas too much, while highway areas consist the major part of the private dataset.
\nOur solution consists of several algorithms, some of them introduced by other competitors and made public, the others developed by ourselves.
\n
The first module is a position estimation based on GNSS data provided in *_derived.csv. With some simple tricks as follow, we achived LB: 6.665 in public.
\nHere we trained a model which estimates the difference between estimated position \\((lat^i_{est}, lng^i_{est})\\) and corresponding ground truth position \\((lat^i_{gt}, lng^i_{gt})\\), based on the provided train data. We introduced several features and used LightGBM for the estimation.
\nBased on the idea introduced here by @dehokanta, we introduced a rule-based outlier detection algorithm. Through several experiments, we've found out that the estimation accuracy improved significantly by rejecting the following points:
\nWe used several types of phones mean method. Extended from the idea introduced here, we used several types of phone means such as:
\nAs discussed in several discussions & notebooks, the noise during the vehicle's stop was apparently one of the major noise in the baseline estimation.
\nWe developed stop detection algorithm using inertial measurements in a rule-based fashion.
For some collections which does not include sensor data, we estimated the stop state only from the position data.
\nWe used snap-to-grid for tree area. For downtown area, we applied a different snap-to-grid method described below.
\nTo overcome some severe problems in simple snap-to-grid, we combined the algorithm with Dynamic Time Warping （DTW）. This algorithm significantly worked especially for the downtown area.
\nWe used a method introduced here by the other team.
\n