Improving accuracy of Pedestrian Detection using Convolutional Neural Networks

Pedestrian detection has become an interesting and challenging problem lately. The main bottleneck in the pedestrian detection is when pedestrians are far away from the camera. The pedestrians who are far from the camera, retaining much less information than the pedestrians who are close, and they also have different appearance. In this article, first a Region-based convolutional neural networks Faster RCNN with a pre-trained Resnet101 architecture is used to provide rich and distinct hierarchical feature representations as well as initial pedestrian proposals. Then, to improve the accuracy of the pedestrian detection, a saliency map is extracted from the video. To do this, the pretrained Yolo is developed to detect the object in the video and the Flownet is applied to extract the motion features. Finally long short-term memory with two convolutional layers is used to obtain the final saliency map. In this research, the Caltech dataset collected by the University of California has been used. The proposed method has been evaluated with a log average miss rate metric, which achieved a miss rate of 40.35%, and compared with the active detection module (ADM) method, 1.92% improvement is achieved.