Towards complex visual surveillance algorithms on smart cameras

One of the drawbacks of currently available smart camera solutions is the gap in comfort and power between software development on the PC and on the embedded system side. This paper describes efforts of bringing high level computer vision software from PCs to the constrained environments of smart cameras. We show that software development with Intel´s Open Source Library OpenCV on smart cameras is feasable and nearly as comfortable as on the PC side, and that processing speeds are sufficiently high for real-time processing. We elaborate on porting efforts of the well known KLT tracking algorithm to a VC4465 smart camera. In addition, we compare the differences in acurracy between the fixed point code and the floating point code of the KLT implementation. Finally, by describing a complex pedestrian tracking algorithm which is suited for crowded scenes, we highlight the practical challenges of running high level vision software on a device with limited computing resources.