Human head detection based on multi-stage CNN with voting strategy

As a classical problem in target detection, human head detection based on head features is an important basis of intelligent vehicle driving and people counting. Due to the irregularity and complexity of human head, artificial designed feature description methods have lower recognition rate and worse robustness. As an important part of deep learning, convolutional neural network (CNN) has applied to image recognition and speech analysis successfully. In view of the instability and hard description of head features, a new head detection method based on multi-stage CNN with voting strategy is proposed in this paper. Firstly, we use features abstracted by multi-stage CNN from different layers to classify respectively. Then, we use the results to get the final classification through voting strategy. Experimental results show that new method has higher recognition rate compared with traditional ones.