CPMS-Net: A lightweight cooperative perception multi-scale network for crowd counting

Abstract Crowd counting aims to estimate the number of people in an image and is a key technology in applications such as intelligent surveillance and public security. With the development of edge computing, deploying crowd counting models on resource-constrained edge devices has become a significant trend, which poses higher requirements for model accuracy, parameter scale, and inference efficiency. However, existing methods universally face a trade-off between performance and complexity: on the one hand, although heavy models possess strong feature representation capabilities, they incur high computational overhead and are difficult to deploy; on the other hand, lightweight models suffer from significant performance degradation in scenarios with complex backgrounds and drastic scale variations due to their limited receptive fields. To address these issues, this paper proposes a lightweight Cooperative Perception Multi-Scale Network (CPMS-Net). Specifically, a Cooperative Perception Module (CPM) is first introduced in the feature extraction stage to simultaneously model inter-channel relationships and spatial positional information. Subsequently, a Multi-Scale Optimization Module (MSOM) is introduced to enhance the model's adaptability to scale variations. Finally, a Lightweight Residual Attention Module (LRAM) is constructed to capture long-range dependencies at a low computational cost. Experimental results on multiple public datasets demonstrate that CPMS-Net, with an ultra-low parameter count of only 0.12M, achieves an MAE of 126.7 and an MSE of 201.0 on the UCF-QNRF dataset. This fully demonstrates its excellent trade-off between accuracy and computational efficiency, validating its application potential for edge device deployment.