CNN implementation of a moving object segmentation approach for real-time video surveillance

In this paper we propose a CNN implementation of an algorithm for moving object segmentation intended for video surveillance applications. The approach is based on the comparison between the current frame and a background dynamically constructed from previous frames. The proposal includes capabilities to detect changes in the illumination conditions as well as to alert against abandoned objects in the control area. The algorithm is composed by simple convolutions and morphological operations as well as simple arithmetic and logic operations which allow the implementation on current focal-plane cellular processor arrays.