Weakly Supervised Crowd Counting via Depth and Density Perception With Dispersed Attention in Smart Surveillance of HMI

Intelligent perception systems in Human-Machine Interfaces (HMI) can significantly enhance user experience. Consumer electronic devices integrated with crowd counting technology, such as security monitoring and smart homes, not only provide real-time environmental awareness but also offer a more convenient user experience. To reduce the computational burden, several weakly supervised crowd counting frameworks have been proposed. However, existing weakly supervised methods ignore the uneven distribution of crowd and the crucial role of depth information in revealing crowd density. In this paper, we propose a novel depth restoration and density perception with cross-dispersed attention network, termed D3CrowdNet, for crowd counting. D3CrowdNet integrates three key modules: i) the depth perception auxiliary module is designed to reconstruct pseudo-3D spatial information of crowds, incorporating spatial distance and scale to enhance the model’s understanding of crowd distribution. ii) We also develop the density-aware graph construction module, which learns interaction relationships between crowd sub-regions. iii) A single branch multi-head cross-dispersed attention mechanism is developed to enable the network to focus on individuals while simultaneously capturing information about the surrounding crowd. Extensive experiments demonstrate that the proposed method achieves superior performance across six benchmark datasets, particularly achieving state-of-the-art results on the ShanghaiTech, UCF-QNRF, and JHU-Crowd++ datasets.