OLODN: An efficient lightweight people detection method for occlusion and crowding scenarios
Wei Sheng, Mingjian Liu, Xiang Li, Mingbao Zhang · IET Image Processing · 2025
Abstract Addressing the real‐time object detection problem for devices with limited computing resources in densely populated and occluded scenarios, a novel occlusion‐aware lightweight object detection network (OLODN) is proposed. This network integrates innovative components to significantly enhance detection efficiency while maintaining high accuracy. Firstly, OLODN employs FasterNet blocks and reparameterised generalised‐FPN to reduce computational complexity and preserve feature extraction and fusion capabilities. Secondly, a reinforced coordination attention mechanism is designed to strengthen the network's ability to capture occlusion boundary information. Additionally, a spatial pyramid pooling feature concatenation module is introduced to integrate multi‐scale features and enhance the algorithm's robustness to occlusions. Lastly, OLODN adopts a task‐aligned one‐stage object detection strategy, optimising the anchor alignment of classification and localisation tasks, effectively improving detection accuracy under occluded conditions. Experiments demonstrate the superiority of the proposed method. On the WiderPerson dataset, OLODN achieved a recall rate of 68.0%, which is 2.8% higher than YOLOv11's 65.2%, while running at 35.3 frames per second (f/s) on CPU and 76.6 f/s on GPU, faster than YOLOv11's 34.1 f/s and 74.5 f/s by 1.2 f/s and 2.1 f/s, respectively.