The problem of video object tracking in a multi-camera scenario, where the cameras and the associated video processing nodes have power constraints, is covered in this paper. The basic application scenario is a grid of identical video processing nodes that wirelessly communicate with each other and a base station. An example of such a system can be used in surveillance systems. Because the power consumption of the individual nodes is an important design criterion, optimizations in the video processing chain or alternative hardware architectures can substantially contribute to an increase in battery lifetime.

Basically, the paper is an extensive study of two different methods for power saving in the described application scenario: the region of interest (ROI) method and the adaptive data partitioning (ADP) method. In the ROI method, a single moving object, bounded by a rectangular region, is processed, ignoring the rest of the frame. The ADP method is totally different. In this method, a node has multiple central processing units (CPUs) and the scene observed by the camera is subdivided among the processors. Only the processors that have a moving scene part are switched on, and the others are switched off. Because of the multiprocessor architecture in the ADP method, each individual processor can operate at a much lower speed compared to the single processor scenario in the ROI method; this results in much lower power consumption per processor. The authors describe a number of possible subframe-to-processor mapping scenarios and formulate a performance model for selecting the optimal number of processors.

Both methods are described clearly and in detail. Various experiments are described to validate and compare the methods. A real system was built for the ROI method, and a multiprocessor simulator was used for the ADP method. The authors conclude that the ROI method results in power savings of 80 percent, compared with 60 percent for the ADP method.

What I like about the paper is the elaborate discussion of all of the decision points and parameter settings involved in selecting the right video object tracking method. Personally, I prefer real experiments because, unfortunately, researchers in academia often rely on simulation only, where the validity of the applied models remains open. This is a somewhat weak point of the paper: the ADP multi-processor experiments are simulation only. I know from personal experience that, in terms of power savings, all system components have to be taken into account to obtain realistic results, something that is not always obvious in a simulation. On the other hand, one could argue that the overall results measured in real systems will generally be worse than those derived from simulations. In this case, the presented simulation results in the paper could be viewed as an upper bound and the conclusions would still be valid.