Research And Application Of Road Target Detection Based On YOLO

Autonomous driving has always been a hot application field of machine learning in recent years,and road target detection is one of the core technologies in assisted driving systems and vehicle automatic driving.Due to the complexity of the actual environment,the detection of road targets has been facing the problems of real-time and poor accuracy.This paper conducts research based on these two practical problems.The main work and innovations are as follows:(1)To solve the problem of slow model detection speed,first recalculate the anchor box parameters of the experimental data set,and then according to the principle of the batch normalization layer,the BN layer parameters of the model YOLOv3-tiny are transformed by mathematical identity Merge into the convolutional layer,calculate the new convolutional layer parameters for training.Finally,a comparative experiment proves that the model detection speed is significantly improved after the parameters are merged.(2)Aiming at the problem of low accuracy of model detection and the problem of model optimization that arises with the increase of the number of network layers,based on the idea of layer jump connection,the dense block is added to the experimental network model.According to the characteristics of the experimental network in this paper,different numbers of dense block are added to the middle and later stages of the network model,and a comparative experiment is carried out with the initial model.The results show that the model added dense block not only improves the detection speed significantly,but also has an advantage over the original model in accuracy.(3)Aiming at the problem that the model is not effective in detecting human targets,the composition of the loss function of YOLO is analyzed,and the GIOU regression loss function is used to replace the original IOU regression loss function in the model.The initial model is compared with the improved model to verify the effectiveness of the improvements.The results show that the overall accuracy of the model has improved slightly,and the accuracy of human targets has increased significantly.(4)A simple road target detection system is implemented,which can detect vehicles,pedestrians and motorcycles in pictures and videos,and supports the comparison of the detection effects of different YOLO models.