18th Place Solution in GSDC

Google Smartphone Decimeter Challengeで18位になりました。
その解法です。
https://www.kaggle.com/c/google-smartphone-decimeter-challenge

Transcript

1. Google Smartphone Decimeter Challenge
   Pre/Post processing (PP)
   Team: JTC
   Private LB: 18th (2.841)
   Reproducing baseline Main process
   OSR data
   from swift-nav
   Calculation for
   satellite position
   (GPS only)
   derived.csv
   reproduced
   baseline location
   Reproduce baseline location by Weighted Least
   Squares(WLS) used satellite position and corrected
   pseudorange
   ・Weight of WLS are hand-tuned.
   ・Use only GPS/GALILEO/QZS
   ・Use scalculated satellite position(GPS only)
   given
   baseline location
   (CV:5.29)
   reproduced
   baseline location v1
   (CV:5.46)
   ① Blending
   baseline location
   GnssLog.txt
   PP
   1. Outlier Correction
   2. Kalman Smoothing
   3. Phone Mean
   4. Remove Phone
   5. Snap to Grid
   6. Phone Mean
   7. Stop Mean
   lightGBM
   PP
   1. Kalman Smoothing
   2. Snap to Grid
   3. Stop Mean
   4. Phone Mean
   IMU data(acce, gyro, magn, rot)
   Apply only
   downtown area
   reproduced
   baseline location v2
   (CV:5.21)
   ×0.5
   ×0.25
   ×0.25
   ③ Estimation for
   location by IMU data
   ② Pre Processing ④ Post Processing
   Weighted
   average
   Area classification
   Copyright 2021 @kuto_bopro
   Apply Kalman Smoothing to IMU
   data and create lag/agg features
   blend
   location
   (CV:4.95)
   Outlier Correction
   ・Judge outlier based on 2σ and
   movement speed > 45m/s
   ・Apply linear interpolation to outlier
   by previous/next position
   Remove Phone
   ・Remove data from a specific
   smartphone and interpolate with
   data from other smartphones.
   Phone Mean
   ・
   ・Mean at the smart phone locations
   in the same collection and epoch
   (epochs were interpolated)
   Snap to Grid
   ・snap prediction points to ground
   truth grids
   ・apply downtown and difficult
   area
   Stop Mean
   ・predict car speed by lightGBM
   ・target: speedMps in ground truth
   ・apply mean for stop groups
   Kalman Smoothing
   ・States:Lat/Lon and epoch
   ・Linear interpolation to keep epoch
   width constant
   lightGBM
   car speed
   grouping if speed < 0.95m/s for
   2s gr1 gr2 gr3
   mean mean mean
   lag/agg feats
   of location
   openstreetmap
   collections
   more 5m
   error points
   highway
   (open sky)
   normal
   downtown
   (high multi path)
   difficult area
   (bridge etc...)
   If the most common
   attribute in the
   "highway" column
   is motorway
   if mean degree >= 3
   in road graph network
   other
   make buffer
   (point to polygon)
   ground truth grids
   prediction points
   

   View Slide