Object detection and localization in compressed video

We study the problem of detecting and localizing objects that are embedded in compressed video sequences. Such a capability has two major and increasingly important practical uses: (1) video surveillance; (2) identification of copyright infringement. We focus here only on the problem of video surveillance. As a general rule, detection and localization of patterns is most efficiently performed in a reduced-dimensional subspace of the original object space. In this regard, it would be ideal to operate directly on the compressed bit stream. As a first step towards doing this, we consider here the problem of detecting and localizing video objects in the DCT domain (i.e., after the quantized DCT coefficients have been decoded but before the inverse DCT has been applied). We present comparisons between this DCT-based approach and the more conventional method in which object detection and localization is performed entirely in the spatial domain.