Real Time ROI Generation for Pedestrian Detection

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. proposed to address this problem. Our method focuses on detecting the Crossing Street Pedestrian (CSP) before a certain distance (5 - 20 meters). Other things will not be cared too much, otherwise system workload will be increased and drivers may be disturbed. The main contribution of this paper can be summarized as: A novel method is proposed to generate ROIs efficiently and keep the number of good ROIs as few as possible. The good ROI has high probability to contain CSP. This will be useful to gain speed of pedestrian detection. The method has the following steps. Firstly, the edges of the CSPs are detected by matching the edge maps of two near frames multiple times after certain proper translating. 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 ROIs are sorted based on their probability values in descending order. At last, HOG and SVM are used to recognize pedestrian in the top 12 ROIs.