International Journal of Computer Theory and Engineering

Abstract

Active object detection (AOD) aims at getting a better image for detection. It moves robot to get a best view of the object. This method is more effective than traditional object detection when facing following problems: the object in image is of tiny scale, the object is blocked by other irrelevant objects, the object is partially captured by camera, etc. We consider that the past images acquired by the robot are related to the problems mentioned above. So different from the state-of-the-art methods which mostly generate actions based on current image, a novel AOD network based on CNN and LSTM networks is proposed to advance detection performance in this paper. The AOD network uses current and historical scenes as well as movements are learned to explore and generate following robot actions. We train the Action Network through reinforcement learning. During training, phenomenon that robot stuck and repetitive in several specific situations usually occurred, which results in invalid training. To solve this, an effective training strategy is proposed to skip trapped or repetitive actions. Our proposed AOD network was evaluated on Active Vision Dataset and the experiment results showed its advantage of both accuracy and efficiency on AOD tasks.