Visual Surveillance using Deep Reinforcement Learning

Abstract

Visual surveillance aims a robust detection of foreground objects, and traditional algorithms usually use a background model image. A current is compared with the background model image. In this paper, we present a visual surveillance algorithm, which determines the parameters in Vibe using deep reinforcement learning. We apply DQN to determine three parameters in Vibe algorithm. We present a policy model which is composed of encoder and decoder type network. Experimental results shows the feasibility of the presented algorithm.