针对在自动驾驶领域中,由于道路场景复杂,目前已有的检测方法存在检测准确率不高,且检测目标单一的问题,提出了一种基于改进YOLOv5s(You Only Look Once version 5)的面向自动驾驶的道路目标检测算法,能够实现车辆、行人、信号灯、交通标志等多个目标的同时检测.首先,在原有模型的基础上,引入EIoU损失函数实现YOLOv5输出端预测框的优化,使收敛速度更快;用YOLOv8的C2f模块替换原模型的C3模块,提高小目标精度;改进YOLOv5的目标检测框架为OTA,在保证检测精度的同时,提升检测速度,降低对设备的要求.然后,在保证以上三者可行的情况下加入单目相机测距,实现目标距离的实时精确跟踪并对危险作出预警.最后,建立数据集并进行数据增强,训练数据集.通过消融试验发现,改进后模型比原模型整体精度提高了 2％,各目标训练精度与召回率的概率均达到99％以上,在目标跟踪实验中能够实时地显示距离并对危险作出预警,证明了该方法是可行和有效的.
Research on Road Target Detection Algorithm and Tracking Based on Improved YOLOv5s
In the field of automatic driving,due to the complex road scene,the detection accuracy is not high for the existing detection methods,and the detection target is single,so an automatic driving-orien-ted road target detection algorithm based on improved YOLOv5s is proposed.It can realize the simulta-neous detection of vehicles,pedestrians,traffic lights,traffic signs and other targets.First,based on the original model,EIoU loss function is introduced to optimize the prediction frame of YOLOv5 output,which makes the convergence speed faster.The C3 module of the original model was partially replaced by the C2f module of YOLOv8 to improve the precision of small targets.The object detection framework of YOLOv5 is improved to OTA,which can improve the detection speed and reduce the requirements on e-quipment while ensuring the detection accuracy.For ensuring the above three feasible,monocular camera ranging is added to achieve real-time accurate tracking of target distance and early warning of danger.Fi-nally,the data set is established and enhanced to train the data set.Through ablation test,the overall ac-curacy of the improved model is 2％higher than that of the original model,and the probability of train-ing accuracy and recall rate of each target is more than 99％.In the target tracking experiment,the dis-tance can be displayed in real time and the danger can be warned,which proves that the method is feasi-ble and effective.
YOLOv5 algorithmautonomous drivingimage recognitionEIoU loss functionC2f moduletarget trackingdanger warning
国家科技期刊平台
NSTL
万方数据
