基于改进YOLOv3的骑车人识别研究

A Cyclist Identification Study Based on Improved YOLOv3

针对智能网联汽车大发展环境下骑车人在公路上为易受伤群体的问题,将目标识别作为无人驾驶技术中的关键一环,提出使用YOLOv3算法对骑车人识别算法进行研究。YOLOv3的主干特征提取网络为Darknet-53,此种网络结构针对于多种类目标检测适用性强,然而公路骑车人作为单种类目标,Darknet-53网络结构显得冗繁。基于此,提出一种在YOLOv3算法基础上记性改进的算法,通过替换主干特征提取网络为Dark-19简化网络结构,降低网络复杂度,之后优化损失函数,将原来的IoU替换成CIoU,以提高识别精度。通过在TDCB上进行仿真实验,结果表明,改进后的YOLOv3算法平均检测精度和检测速度都有所提高,精度上提高了约3%,检测速度上约提高了0.013 s,此种改进后的算法有助于提高公路骑车人的安全性,对骑车人识别研究有着重要意义。

Aiming at the problem that cyclists are vulnerable groups on the road under the development environment of intelligent network-connected automobile,target recognition is also regarded as a key link in driverless technology,and it is proposed to use YOLOv3 algorithm to study the cyclist identification algorithm.The backbone feature extraction network of YOLOv3 is Darknet-53,which is suitable for multi-species target detection,but the network structure of Darknet-53 network appears redundant.An algorithm based on the YOLOv3 algorithm is proposed to simplify the network structure and reduce the complexity of the network by replacing the trunk feature extraction network for Dark-19,and then optimize the loss function and replace the original IoU with the CIoU to improve the recognition accuracy.Through simulation experiments on TDCB,the results show that the improved YOLOv3 algorithm has improved the average detection accuracy by about 3%,and the detection speed by reducing about 0.013 s.