Learning-Deep-Learning

Monocular Velocity: Camera-based vehicle velocity estimation from monocular video

July 2020

tl;dr: Relative velocity estimation from a sequence of monocular images, taken with a moving camera.

Overall impression

This is the winning entry to the monocular velocity estimation challenge. Lightweight trajectory based features (list of bbox location) are good enough. Better than full solution with depth and optical flow features.

The SOTA error is around 1.12 m/s, as compared to the GT error of 0.71 m/s.

Key ideas

Technical details

Notes