Crowd Counting via Enhanced Feature Channel Convolutional Neural Network

Abstract

Accurate crowd counting is important for interpreting and understanding crowd, which has great practical significance in video monitoring, public safety, urban planning, the construction of intelligent shopping malls and so on. For accurate counting, many excellent algorithms have been proposed, but there are still some challenges in terms of scale variation, occlusion, inaccurate counting in various backgrounds and so forth. In this paper, we propose a new model EFCCNN (Enhanced Feature Channel Convolutional Neural Network) to deal with these challenges. The proposed EFCCNN model has three main contributions. We propose a new convolutional neural network, which can be trained by end-to-end, and it performs better than other crowd counting networks. Additionally, we use SENet (Squeeze-and-Excitation Network) structure to change the channel weight, which can enhance the significant channel, and we use residual structure to transmit the channel weight to improve the counting precision. The SENet structure is helpful to solve the problem of scale variation and occlusion. The EFCCNN model is the first crowd counting model using channel weight information. Furthermore, a new loss function focusing on the structural information of images is proposed, which reduces the mean absolute error of crowd counting, effectively solves the problem of inaccurate counting in various backgrounds, such as crowd miscounting in the tree and brush background, and improves the quality of the crowd density map on SSIM (Structural Similarity Index Measure). Experiments on ShanghaiTech, Mall, UCF_CC_50 dataset show that EFCCNN have a lower mean absolute error of crowd counting and a higher quality density map.