A novel memory mechanism for video object detection from indoor mobile robots

Video object detection has great potential to enhance visual perception abilities for indoor mobile robots in various regions. In this paper, a novel memory mechanism is proposed to enhance the detection performance for moving sensor videos (MSV), which obtain from indoor mobile robot. And the proposed mechanism could be applied as an extension module for a number of existing image object detectors. First, we analyze characteristics of the indoor MSVs, concluding the key characteristics as mild changes, complicated contents and relative movements. Second, a memory-unit dispatching and application method is devised to maintain prior memory contents and utilize the contents to achieve better detection performance. Finally, we create a corresponding indoor MSV dataset and compress the mechanism into a module to evaluate its localization performance. Our experiment results are presented to illustrate the proposed mechanism and achieve an average localization margin by 19.8% compared with several representative original detectors.