人员密集场景的目标检测在实时系统中至关重要,但面临硬件资源有限和遮挡问题,导致检测延迟和精度下降。提出了一种遮挡感知轻量级目标检测方法,包括主干、特征融合和输出预测3部分。该方法使用快速网络块提取特征,并通过位置注意力机制关注遮挡边界。主干部分的特征金字塔串联汇聚模块减少信息丢失,提高对不同尺度和遮挡人员的识别能力。特征融合部分采用分组洗牌卷积,优化特征流动而不增加计算负担。输出预测部分使用任务对齐单阶段目标检测方法,提升遮挡条件下的识别准确性。实验结果显示,所提方法在WiderPerson数据集上的召回率达66.8%,比YOLOv8-n高2.0个百分点,且模型参数量仅1.8×106,运行效率优于其他模型。在UpDown数据集上,分类错误率和未检测目标错误率分别为2.6%和1.3%,分别比YOLOv8-n的低了0.4个百分点和0.7个百分点。实验验证了该方法在资源有限设备中的高效性。

Object detection in crowded scenarios is crucial in real-time systems, but it faces chal- lenges such as limited hardware resources and occlusion issues, leading to detection delays and reduced accuracy. This paper proposes an occlusion-aware lightweight object detection method (OLODN) comprising 3 parts: a backbone, feature fusion, and output prediction. The method employs fast network blocks for feature extraction and utilizes a positional attention mechanism to focus on occlusion boundaries. The spatial pyramid pooling feature concatenation module in the backbone reduces information loss and enhances the ability to recognize individuals of varying scales and occlusions. The feature fusion section adopts grouped shuffle convolution to optimize feature flow without increasing computational overhead. The output prediction section employs a task-aligned single-stage object detection method to improve recognition accuracy under occlusion conditions. Experimental results show that the method achieves  66.8% recall on the WiderPerson dataset, which is 2.0 percentage points higher than that of YOLOv8-n, with only 1.8×106 model parameters and superior operational efficiency compared to other models. On the Up-Down dataset, the classification error rate and undetected object error rate are 2.6% and 1.3%, respectively, which are 0.4 percentage points and 0.7 percentage points lower than YOLOv8-n. The experiments validate the methods efficiency on resource-constrained devices.