Running person detection from a community patrol robot

Abstract

In this paper, a running person detection method is proposed for the community patrol robot. The challenges include the diversity of movement direction and the ego-motion of camera. The diversity of movement direction means that it is difficult to gain high accuracy detection by only using appearance information. The ego-motion of camera means that the motion information contains high noise. To address these limitations, two-stream architecture of convolutional networks, spatial stream and motion stream, is proposed to capture the complementary information on appearance from still frames and motion between frames. In addition, simple but effective filtering based ego-motion elimination technique is applied in motion stream input calculation. We demonstrate the efficiency of the proposed method on the video captured from the community scene and the performance compared with existing methods, which validates that the proposed method detects running person more accurately.