Real-time running detection from a moving camera
Qingtian Wu, Huiwen Guo, Xinyu Wu, Shibo Cai, Tao He, Wei Feng · 2016
In this paper, we present a real-time running detection system from a moving camera. 11 fps and satisfying detection accuracy in outdoor surveillance environment can be achieved in the system, using only one processing thread without resorting to special hardware like GPU. Real-time and high accuracy detection are made possible by two contributions. First, we use a succession of preprocessing methods to extract regions of interest (ROI), including spatial domain analysis, computing the optical flow of every two consecutive images and predefining a threshold based on optical flow to choose effective areas as ROIs. Second, we use relative small-batch samples to train our 5-layer CNN in order to achieve a satisfying detection rate. Experiments on various videos shot in different time and places have proved that the proposed system can detect running pedestrians in outdoors effectively and robustly and meet the real-time requirement with a high detection rate.