Spatial object detection and classification in JPEG bitstreams

To reduce storage and transmission requirements, digital images are generally compressed in some fashion. Consequently, if one is interested in detecting and classifying spatial objects of interest within an image, it might, in many instances, be more efficient to do so in the compressed domain because less data would need to be processed and the computation required to decode the image would be avoided. In our earlier work, we have shown that object detection in the JPEG bitstream domain is both effective and efficient. In this paper, we expand on this earlier work, addressing issues such as detection within a constant false alarm rate (CFAR) context, detection over multiple frames, and multiple objects classification - all within the bitstream of a JPEG-compressed image.