Pedestrian Detection Algorithm based on Improved Faster RCNN

Abstract

To solve the problem of poor detection effect of small-scale pedestrian in the image, based on the existing deep convolutional neural network(CNN)model. We propose an improved pedestrian detection algorithm based on Faster RCNN. First, the image features are extracted by CNN, and the areas that may contain pedestrians are extracted by clustering and regional Recommendation Network (RPN), Secondly, a multi-layer feature fusion strategy based on cascade is proposed. The semantic information of the network can be enhanced by combining the features of the high level with those of the low level. Finally, the Online Hard Example Mining(OHEM) method is used to train the high loss samples to deal with the imbalance between the positive and negative samples, so as to significantly improve the detection performance of the algorithm. Experimental results show that this method greatly improves the accuracy of small target pedestrian detection. In PASCAL VOC 2007 and INRIA pedestrian data sets, the average accuracy was improved by 6.3% and 13.93% respectively.