基于改进YOLOv5的拥挤行人检测算法

Crowded Pedestrian Detection Algorithm Based on Improved YOLOv5

针对密集场景下行人检测的目标重叠和尺寸偏小等问题,提出了基于改进YOLOv5的拥挤行人检测算法。在主干网络中嵌入坐标注意力机制,提高模型对目标的精准定位能力;在原算法三尺度检测的基础上增加浅层检测尺度,增强小尺寸目标的检测效果;将部分普通卷积替换为深度可分离卷积,在不影响模型精度的前提下减少模型的计算量和参数量;优化边界框回归损失函数,提升模型精度和加快收敛速度。实验结果表明,与原始的YOLOv5算法相比,改进后YOLOv5算法的平均精度均值提升了7.4个百分点,检测速度达到了56.1 f/s(帧/秒),可以满足密集场景下拥挤行人的实时检测需求。

Aiming at the problems of mutual occlusion and small target size in pedestrian detection of dense scenes,a crowded pedestrian detection algorithm based on improved YOLOv5 was proposed.Firstly,the coordinate attention mechanism in the backbone network was embed to enhance the accurate positioning ability of the model to the target.Secondly,on the basis of the original algorithm's three-scale detection,the shallow detection scale was added to improve the detection effect of small sized targets.Thirdly,the depth separable convolution was used to replace some ordinary convolution,which cowld reduce the calculation and parameters of the model without affecting the accuracy of the model.Finally,the bounding box regression loss function was optimized to improve the model accuracy and speed up the convergence speed of the model.Experiments show that,compared with the original YOLOv5 algorithm,the average accuracy of the improved YOLOv5 algorithm has increased by 7.4 percentage points,and the detection speed has reached 56.1 frames/s,which can meet the real-time detection requirements of crowded pedestrians in dense scenes.