MVmed: Fast Multi-Object Tracking in the Compressed Domain
Lukas Bommes, Xinlin Lin, Junhong Zhou
Abstract
We present MVmed, an algorithm for real-time online tracking of people and objects in MPEG-4 and H.264 compressed videos and integrate it into a multi-purpose tracking software for manufacturing sites. To support arbitrary video sources with no prior setup our tracker needs to be compatible with a variety of video codecs and camera settings. Existing compressed domain trackers are limited in this regard. They require a fixed interval of key frames, use only P frames and usually support only a single codec. MVmed overcomes these limitations and supports both MPEG-4 and H.264 codecs, P and B frames and arbitrary key frame intervals. On the MOT17 benchmark MVmed achieves a MOTA of 45.3 % at 42.1 Hz (266.9 Hz without detection) which is as accurate but significantly faster than the previous state of the art in compressed domain tracking. With this work we release the source code of MVmed and a Python package for motion vector extraction from video.