Robust object detection using cascade filter in MPEG videos

Abstract

We propose a novel approach for motion vector (MV) based object detection in MPEG-2 video streams. Rather than processing the extracted MV fields that are directly extracted from MPEG-2 video streams in the compressed domain, we perform MV smoothing, perform MV noise reduction, obtain more robust object information, and refine this information through a cascaded filter composed of a Gaussian filter and a median filter. As a result, the object detection algorithm is more capable of accurately detecting objects. We compare the performance of our proposed system with the popular and commonly used spatial filter processing techniques: median filter, mean filter, Gaussian filter, and no filter. Based on experimental results performed over the MPEG7 testing dataset and measuring performance using the standard recall and precision metrics, object detection using the cascade filter is remarkably superior to the alternative filtering techniques. In addition to these results, we describe a user system interface that we developed, where users can maintain the filter parameters interactively.