Arbitrary Perspective Crowd Counting via Multi Convolutional Kernels

Abstract

Cross-scene crowd counting plays a more and more important role in intelligent scene monitoring, and it is very important in the safety of personnel and the scene scheduling. The traditional estimation of crowd counting is mainly dependent on the simple background of scenes, which is not conducive to the complex background. To address this problem, in this paper, we propose a multi convolutional kernels net for crowd counting, which discards the subjectivity and the occasionality of the traditional manual feature extraction. Firstly, we label dataset for convolution output features. Then we use the fully convolutional network to create the density map at the end of the network with multi convolutional kernels. Finally, we perform integral regression on density maps to estimate the crowd counting. The dataset that we used is a set of publicly available datasets, which are the Shanghaitech dataset, the UCF_CC_50 dataset and the UCSD dataset. The experiments based on video images show that the proposed method is more effective than traditional methods in terms of robustness and accuracy.