Drone2tiles mapping

As part of work, road semantic segmentation has been developed.

Demo: YouTube

todo:

• test on more videos
• investigate problem with higher frame video input
• improve template matching of drone position
• add forward intertial kinematics to kalman fitler to take into account velocity (add velocity to state vector)
• convert points to real GPS locations
• fetch gps tile automatically based on current location
• create tool for gps tile download

Release history

Version 0.0.5

Added hard-coded visualization for road crossing on satellite tile. Must be removed and replaced with automatic tile fetcher and gps to image coordinates translation.

Version 0.0.4

Improved road crossing detection and visualisation by combining results from contour recognition and hough lines intersection.

Version 0.0.3

Added tracking of road crossing on drone view

Version 0.0.2

Added Kalman filter, faulty prediction filtering based on distance threshold.

Version 0.0.1

Canny edge preprocessing, scale invariant template matching, each frame independently.

ffmpeg commands

Create video from images:

ffmpeg -framerate 30 -pattern_type glob -i '*.jpeg' -c:v libx264 -pix_fmt yuv420p out.mp4

If error height is not divisible by two - add -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2".

ffmpeg -framerate 10 -pattern_type glob -i '*.jpeg' -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2" -c:v libx264 -pix_fmt yuv420p ../videos/0.0.3.mp4

ffmpeg -i ../DJI_0012.MP4 -vf fps=30 out%d.jpeg