Temporal Motion Vector Filter for Fast Object Detection on Compressed Video

A novel Temporal Motion Vector Filter (TF) is presented and evaluated for real-time object detection on compressed videos in MPEG-2, MPEG-4 or H.264/AVC formats. The filter significantly reduces the noisy motion vectors that do not represent a real object movement . The filter analyses the temporal coherence of block motion vectors to determine if they are likely to represent true motion in the recorded scene. Experiments are performed using the CLEAR metrics for object detection and public available video datasets from CAVIAR, PETS and CLEAR. These experiments demonstrate that the TF outperforms the Vector Median Filter, by providing better object detection accuracy with reduced computational complexity. The good results obtained by the TF make it suitable as a first step towards implementing systems that aim to detect and track objects from compressed video by using motion vectors. The TF could also be used to improve other techniques based on motion vectors such as Global Motion Estimation (GME) and Motion-Compensated Frame Interpolation (MCFI).