Block compressed sensing based background subtraction for embedded smart camera

Embedded smart camera networks represent an emerging direction of next generation surveillance systems. A big challenge to implement computer vision applications on embedded cameras is the limit of memory and computational capacity. Since background subtraction algorithms play a fundamental yet significant role of most computer vision applications, their memory requirements and computational efficiency should be taken into account in the design. In this paper, we propose an efficient hierarchical light-weight background subtraction approach by combining the pixel-level and the block-level background subtraction modules into a single framework so that it is capable of dealing with dynamic background scenes. Block compressed sensing theory is for the block-level module design to save memory and improve computational efficiency. Moreover, considering the continuity of foreground objects, a novel integral filter is designed for the pixel-level module to eliminate perturbations efficiently. Experimental results on various videos demonstrate superior performance of the proposed algorithm. The proposed light-weight algorithm only requires about 6.5 bytes per pixel, and is applicable for embedded smart cameras. Furthermore, as each block is processed independently, it can be implemented in parallel.