MPANet: A Multi-stage Pixel-Level Attention Network for Crowd Counting

Abstract

Crowd counting is widely used in public security, video surveillance, smart city construction and other fields, and has important and positive significance for controlling the number of people in specific places, directing public transportation, and ensuring social stability. The traditional counting method has low accuracy, limited scene and huge volume. With the development of deep learning, the traditional method is gradually replaced by the convolutional neural network method. We built a crowd density estimation algorithm based on convolutional neural network with high intelligence and low power consumption, which takes into account the accuracy of crowd counting and the reasoning speed. Compared with the existing algorithms, our algorithm incorporates pixel-level attention mechanism at different stages, so that the neural network can pay attention to useful features, so as to improve the accuracy of crowd counting. Secondly, most existing methods based on pixel-level attention mechanism use additional networks to extract information, while the attention features and density maps of our network are generated in a unified network, which improves the reasoning speed of the model.