A Performance Comparison of Pedestrian Detection Using Faster RCNN and ACF

Abstract

This paper presents a performance comparison of pedestrian detection using faster RCNN and ACF. RCNN finds independently candidate areas for detectable object and extracts feature vector of CNN from each regional image. These images are classified by class-specific linear SVMs. ACF computes several channels from one image and gets a low resolution channel by integrating them into smoothed one. Each pixel becomes a feature and these pixels turn a feature vector. Finally, pedestrian and background are separated by using decision tree and boost. The databases used in this study are downloaded from Youtube. The videos include vehicle driving environment scenes. The experiment result shows that precision of faster RCNN is 56.73% higher than precision of ACF on the manual works. To automating works, label image is formed by human. And recall rate and precision are compared. The ACF detector with Caltech model shows good recall rate than other methods but its precision is the worst. The faster RCNNs precision is 7 times higher than second better method.