Robust Compressed Domain Object Detection in MPEG Videos

A.M.A. Ahmad, D.-Y. Chen, and S.-Y. Lee (Taiwan)


MPEG2, Filter, Object Detection, MV, Gaussian


In this paper we propose a novel approach for Motion Vector (MV) based object detection in MPEG-2 video streams. Instead of designing an object detection algorithm by processing the extracted MV fields that are directly extracted from MPEG-2 video streams in the compressed domain, we smooth the MVs which contain magnitude and direction information, eliminate the noise within their contents, obtain more robust object information and refine it through a Gaussian filter. Then, the object detection algorithm will be more capable of detecting objects based on the filtered MV magnitude and direction correctly and meaningfully. We compare the object detection performance among using three popular and commonly used spatial filters: median, mean and Gaussian filter. Based on experimental results over MPEG7 testing dataset, object detection using Gaussian filter performance based on the standard recall and precision as metrics is remarkable and strongly effective over other filters as well as over no filter. We develop user system interface, where users can maintain the filter parameters interactively.

Important Links:

Go Back