Compressed domain indexing of scalable H.264/SVC streams

We present methods to efficiently analyze scalable, compressed H.264/scalable video coding (SVC) video streams. Relying solely on information present in the compressed stream, we estimate the global camera motion, perform motion segmentation and use a simple matching process to track moving objects over time. Object energy images are constructed in order to help resolve the problem of object correspondence during the occlusions of multiple objects. To save computing time, we analyze lower spatial layers of the stream and add higher layer information only if necessary. We draw 2-D object trajectories in the view plane of the camera and use the temporal evolution of the objects' properties to estimate the relative distance to the camera, resulting in a pseudo 3-D representation of the trajectories. Finally, the suitability of the motion parameters to perform video retrieval/copy detection tasks is demonstrated. We therefore form two simple descriptors that are invariant to a series of transformations.