Real-time Steal Recognition on CCTV-Based Videos for Embedded Systems

Abstract

Action Recognition is a computer vision task in which a given video has to be classified. As far as videos should be processed, this task is computationally more expensive than the other common tasks of computer vision such as classification and object detection. There will be many issues to address when this should be implemented, such as how to handle the computational costs of this task while working in a real-time manner, especially when it is being conducted on embedded devices as well. This paper explores surveillance as one of the situations in which action recognition becomes so critical. A CNN and RNN-based solution have been introduced. Then some experiments have been conducted in order to determine the best architecture choice for each of the CNN and RNN parts. As a result, the final can be used on embedded devices real time maintaining high accuracies.