An effective approach to crowd counting with CNN-based statistical features
Shunqiang Liu, Sulan Zhai, Chenglong Li, Jin Tang · 2017
Recent works on crowd counting have achieved promising performance by employing the Convolutional Neurol Network (CNN) based features. These works usually design a deep network to detect pedestrian heads, and then count them. In this paper, we propose a novel approach to count pedestrians effectively based on the statistical CNN features. In particular, our approach only uses the first layer features of the CNN pre-trained offline on ImageNet, and thus obtains an efficient solution for crowd counting. Then, by analyzing the statistical properties of the first layer features, we observate the number of people fluctuates according to the value of the statistical features. Therefore, we employ these statistical features to train SVM, and can thus directly obtain the number of pedestrians. Experimental results on standard benchmark, UCSD, verify the effectiveness of the proposed approach.