| 康比比,董小明.改进YOLOv10的密集人群检测算法研究[J].电子测量与仪器学报,2025,39(12):248-257 |
| 改进YOLOv10的密集人群检测算法研究 |
| Pedestrian detection algorithm in dense scenes based onimproved YOLOv10 algorithm |
| |
| DOI： |
| 中文关键词: YOLOv10 SIoU 密集人群检测 CBAM DySample |
| 英文关键词:YOLOv10 SIoU intensive crowd detection CBAM DySample |
| 基金项目:安徽省优秀拔尖人才项目（gxgwfx2020053）资助 |
|
|
|
|
| 摘要点击次数: 618 |
| 全文下载次数: 449 |
| 中文摘要: |
| 密集人群检测算法在公共安全监控、智能交通调度等领域意义重大。针对密集场景中目标遮挡、小目标检测精度低及漏检问题,以轻量化 YOLOv10 为基础,提出改进的 YOLOv10-SCD 算法。首先，融入卷积块注意力模块(convolutional block attention module,CBAM),通过通道 - 空间双维度加权,增强运动模糊处理能力并提升行人检测精度;其次,引入动态采样(dynamic sample,DySample)上采样器提升图像分辨率与处理效率,同时采用改进交并比(SIoU)优化损失函数,进一步提升定位精度与边界框回归速度;最后,在密集人群数据集上验证算法性能,并通过消融实验分析各模块作用。实验表明,YOLOv10-SCD 相较于原始 YOLOv10,核心指标显著提升，其中精确度提升 1.5%、召回率提升2.9%，平均精度均值（mAP）mAP@0.5提升0.8%、mAP@0.5:0.95提升1.8%等。消融实验在两套数据集进行，自建数据集聚焦分析各模块对算法精度、效率的单独与协同影响;WiderPerson 公开数据集则验证算法泛化能力。因此,YOLOv10-SCD 能高效应对密集人群复杂场景,缓解目标遮挡、小目标难识别问题,显著提升目标检测鲁棒性与综合性能。 |
| 英文摘要: |
| The dense crowd detection algorithm is of great significance in fields such as public safety monitoring and intelligent traffic scheduling. Aiming at the problems of target occlusion, low detection accuracy of small targets, and missed detection in dense scenes, this paper, based on the lightweight YOLOv10, proposes an improved YOLOv10-SCD algorithm. First, the convolutional block attention module （CBAM) is integrated. Through channel-space two-dimensional weighting, the motion blur processing ability is enhanced and the pedestrian detection accuracy is improved. Second, the dynamic sample (DySample) up-sampler is introduced to improve the image resolution and processing efficiency. At the same time, the SIoU optimization loss function is adopted to further improve the positioning accuracy and the bounding box regression speed. Finally, the performance of the algorithm is verified on the dense crowd dataset, and the role of each module is analyzed through ablation experiments. Experiments show that compared with the original YOLOv10, the core indicators of YOLOv10-SCD are significantly improved:the precision is increased by 1.5%, the recall rate is increased by 2.9%, mAP@0.5 is increased by 0.8%, and mAP@0.5:0.95 is increased by 1.8%. The ablation experiments are carried out on two sets of datasets,the self-built dataset focuses on analyzing the individual and synergistic effects of each module on the algorithm accuracy and efficiency; the WiderPerson public dataset verifies the generalization ability of the algorithm. Therefore, YOLOv10-SCD can efficiently cope with the complex scenes of dense crowds, alleviate the problems of target occlusion and difficult recognition of small targets, and significantly improve the robustness and comprehensive performance of target detection. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
