关键词: 人员密集检测, 人员行为遮挡检测, 计算资源受限, 类间遮挡和类内遮挡, 增强位置注意力机制模块, 特征金字塔串联汇聚模块

Abstract: Target detection in crowded scenes is crucial for real-time systems but faces challenges such as limited hardware resources and occlusion issues, leading to delays and reduced accuracy. This paper proposes an Occlusion-aware Lightweight Object Detection Network(OLODN), consisting of three components: Backbone, Neck, and Head. The network uses FasterNet Block for feature extraction and employs Reinforced Coordination Attention to focus on occlusion boundaries. The Backbone incorporates Spatial Pyramid Pooling Feature Concatenation to minimize information loss and enhance recognition of various scales and occluded objects. The Neck uses Grouped Shuffle Convolution to improve feature flow and integration without additional computational burden. The Head adopts Task-aligned One-stage Object Detection to enhance object recognition accuracy under occlusion conditions. Experimental results show that OLODN achieves a recall rate of 66.8% on the WiderPerson dataset, 2.0% higher than YOLOv8-n, with a model size of only 1.8M, and performs more efficiently than other models. On the UpDown dataset, the classification error rate and missed detection rate are 2.6% and 1.3%, respectively, lower than YOLOv8’s 0.4% and 0.7%. The experiments confirm the efficiency of OLODN in handling occlusion issues on resource-constrained devices.

Key words: Crowd Detection, Occlusion Detection in Human Behavior, Resource-constrained Computing, Inter-class and Intra-class Occlusion, RCA Attention Mechanism, SPPFC Module.