| 摘 要: 针对现有的行人检测算法在复杂场景下检测速度慢、检测精度不高的问题,提出一种轻量化密集行人检测算法MER-YOLO(Miniature Enhanced Recognition-You Only Look Once)。首先,MER-YOLO以MobileNetV3(轻量化网络模型)作为特征提取主干网络,提升模型对于小目标及模糊图像的学习能力;其次,通过融合深度可分离卷积和ECA(Efficient Channel Attention)注意力机制构建DPE-C3模块,解决密集行人检测过程中的遮挡丢失漏检的问题;最后,MER-YOLO使用空间和通道重建卷积处理标准卷积中固有的空间和信道冗余,减少模型计算需求。该算法应用于WiderPerson(混合行人数据集)上的检测精度达到了78.9%,相较于YOLOv5s算法提升了3.0百分点,同时模型计算量比YOLOv5s降低了13.3百分点。因此,MER-YOLO算法兼顾了检测准确度和检测速度的要求。 |
| 关键词: 行人检测;轻量化网络;注意力机制;空间重建卷积 |
|
中图分类号: TP391.4
文献标识码: A
|
|
| Lightweight Dense Pedestrian Detection Algorithm Based on MobileNet |
|
WEI Zhi1, LIU Gang1,2, ZHANG Xu1
|
(1.School of Electronic and In f ormation Engineering, Nanjing University of In f ormation Science and Technology, Nanjing 210044, China;
2.Wuxi University, Wuxi 214105, China)
1542975776@qq.com; liugang@cwxu.edu.cn; 1793697124@qq.com
|
| Abstract: To address the issues of slow detection speed and low detection accuracy of existing pedestrian detection algorithms in complex scenes, this paper proposes a lightweight dense pedestrian detection algorithm called MER-YOLO (Miniature Enhanced Recognition-You Only Look Once). Firstly, MER-YOLO utilizes MobileNetV3 as the feature extraction backbone network to enhance the model's learning ability for small targets and blurry images. Secondly, by integrating depthwise separable convolution and ECA (Efficient Channel Attention) attention mechanism to construct the DPE-C3 module, the algorithm tackles problems of occlusion, loss, and missed detection in dense pedestrian detection. Lastly, MER-YOLO employs spatial and channel reconfiguration convolution to handle inherent spatial and channel redundancies in standard convolutions, reducing the computational requirements of the model. The proposed algorithm achieves a detection accuracy of 78.9% on the WiderPerson dataset (a mixed pedestrian dataset), which is 3.0 percentage points higher than the YOLOv5s algorithm. At the same time, the number of parameters is reduced by 13.3 percentage points compared to YOLOv5s. Therefore, the MER-YOLO algorithm meet the requirements of detection accuracy and detection speed. |
| Keywords: pedestrian detection; lightweight network; attention mechanism; spatial reconfiguration convolution |
