Real Time ROI Generation for Pedestrian Detection

Xin Zhao,Mao Ye,Yingying Zhu,Chuanzhi Zhong
DOI: https://doi.org/10.1109/CISE.2009.5366155
2009-01-01
Abstract:Efficient pedestrian detection is essential for intelligent vehicles and driver assistance system. An increasing number of experts have attached more importance to this subject in recent years. The input of this task is a video captured by a monocular optical camera which is installed on a vehicle. And the aim is to locate every pedestrian in each frame of the video as soon as possible. This task has two steps, i.e., region of interest (ROI) generation and pedestrian recognition in the ROIs. Since most of previous methods face the issue of large number of ROIs, so real time requirement is difficult to be achieved. In this paper, a novel method is proposed to address this problem. Our method can generate ROIs efficiently and keep the number of good ROIs as few as possible. Good ROI means it has high probability to contain crossing street pedestrian (CSP), which should be particularly concerned by drives. Our method has the following steps. Firstly, the edges of the CSPs are detected by matching the edge maps of couple frames multiple times. Secondly, ROIs are generated based on the detected CSP edges and pedestrian shape features. Thirdly, the probability of each ROI containing CSP is calculated. And these ROIs are sorted by their probability values in descending order. At last, HOG and SVM are used to recognize pedestrian only in the top 12 ROIs. The results from real experiments confirm the good performance and high efficiency of our method.
What problem does this paper attempt to address?