基于YOLOv5网络的径赛人体目标检测

Human Target Detection in Track Events Based on YOLOv5 Network

针对径赛人体目标识别中,在图像上重叠、粘连的运动员检测精度低,容易漏检并在多目标识别中检测速度较慢的情况,本文提出一种基于YOLOv5的径赛人体目标检测模型。首先,在专业场地拍摄运动员冲刺影像,对所有运动员与遮挡目标进行标注,制作成专用径赛识别数据集。其次,通过数据增强的手段,将数据集中运动员部分身体进行遮挡,获取大量的差异性增强数据,提高模型对遮挡目标识别度。最后,对YOLOv5网络进行优化,将网络中Leaky Relu激活函数与Hardswish激活函数全部改为Silu激活函数,简化了体系结构;改进nms非极大值抑制机制,增加模型对重叠目标的敏感度。实验结果表明,改进后的YOLOv5模型对重叠的人物有着更为精确的检测效果。模型平均精度mAP@0.5值为0.982,测试精度为0.969,单张图片处理的平均时间为0.012 s,目标检测速度达到83.3FPS。能够满足径赛检测的需求。

Aiming at the problems of low detection accuracy,easy miss detection and slowly process in trace events which cause by athletes overlap in the picture. A new method of human target detection based on improved YOLOv5(You Only Look Once version5)was proposed in this paper.First of all,to build a data set for trace events. More than 2000 pictures was taken in the professional playground. Secondly,with data enhancement of random occlusion.The data set get double and much more difference enhancement data was included which make the data set more sensitive to obscure targets. Last but not least. Changed the activation function into SiLU can simplify the YOLOv5 networks.Rewrite the NMS(Non-Maximum Suppression)mechanism to make the network more interest in overlap targets.The experimental results show that the new YOLOv5 model can better work in trace events of human detection.The accuracy of the model increases to 0.969 and the mAP@0.5% increases to 0.982. And 83.3FPS can meet the trace needs.