Pedestrian Detection Using ACF Based Fast R-CNN

Accurate and efficient performance is an important requirement for pedestrian detection. In this paper, we propose a novel detection framework named as ACF Based Fast R-CNN (ABF-CNN). The ABF-CNN consists of a ACF proposal generation part and a Fast R-CNN detection network. The motivation to use the Aggregated Channel Features (ACF) is due to its real-time efficiency and effective performance. To achieve high accuracy, we further propose to make use of the deep learning method Fast-RCNN. Furthermore, in order to solve the problem that CNN based methods have difficulty in hard negative mining, we propose to integrate Online Hard Example Mining (OHEM) training strategy into our detection framework. By thoroughly analyzing and optimizing each step of pedestrian detection pipeline, we develop an accurate detection framework with low computational complexity. The experimental results demonstrate that our framework achieves state-of-the-art performance on Caltech pedestrian dataset with 17% miss rate.