Human Action Recognition from Video Bitstreams

12/2014 - 05/2015

Project:

Directly using the Motion Vectors (MV) from the compressed MPEG-4 video bitstreams to achieve fast Human Action Recognition, bypassing video decoding process and avoiding pixel-domain analysis.

  • Extract MV-based Spatial-temporal Interest Points from video bitstream instead of from the decoded video
  • Leverage MV motion features for classification–using traditional BoW and Fisher Vector followed by SVM as the framework.
  • C++
  • OpenCV
  • FFMPEG
  • Linux