TFA-CNN: an efficient method for dealing with crowding and noise problems in crowd counting

Crowd counting technology is to let people understand the spatial distribution of crowds in various scenes. In reality, a large number of occlusions and scale variations make it extremely challenging to achieve accurate counting in crowded venues. Aiming at these problems, this paper designs a crowd density estimation network that can maintain good accuracy in scenes that are both crowded and have large-scale changes: Texture Feature Attention Convolutional Neural Network (TFA-CNN). Specifically: (1) A Differential Texture Module (DT Module) is proposed to identify various texture features of the bottom feature map and to better distinguish between background and foreground regions; (2) proposed the Multi-Channel Threshold Replacement Attention Module (MTRA Module), which combines channel and spatial attention mechanisms to allow the network to pay more focus on the head position of the crowd, thereby reducing the counting error. TFA-CNN has conducted multiple experiments on several publicly available and challenging datasets, and the results are superior to many SOTA methods, demonstrating excellent generalization and robustness.