Research On Real-time Vehicle Detection Algorithm Based On YOLO Model

In recent years,deep learning develops rapidly.Vehicle real-time detection uses more deep learning related technologies to complete real-time detection tasks.Research on vehicle real-time detection technology can detect vehicles in front of the road and make safety warning.Because the actual traffic scene is complex and changeable,and the vehicles on the road have different models,colors,brightness,occlusion and other factors,the traditional vehicle real-time detection algorithm can not meet the actual needs,so it is a challenging task to study the vehicle real-time detection algorithm with better indicators in all aspects.YOLO model not only overcomes the disadvantage of slow speed of traditional vehicle real-time detection algorithm,but also has better detection accuracy,which provides a new way for real-time detection of road vehicles.Therefore,this dissertation studies the real-time vehicle detection algorithm based on the YOLO model.The main research contents are as follows:In order to effectively verify and train the designed model algorithm,the vehicle data set is self-made and expanded.When collecting the data set,it takes pictures from different scenes,mainly including various weather conditions,different light and shade conditions,whether there is target occlusion or not,and the distance between the vehicle and the acquisition equipment.In view of the problem that the current vehicle real-time detection algorithm is difficult to meet the requirements in terms of detection frame rate,the framework structure of the YOLO model is studied in depth by adding the darknet-53 classification network to the YOLO v3 model,the extracted features can be classified more quickly and accurately,and the network structure parameters can be reset in view of the situation that the vehicle driving speed is fast and the target background is complex.In the model training stage,the loss function is studied,and the mean square and error loss function are used to train the model to the optimal state.The experimental results show that the performance of the YOLO model is better than that of Fast RCNN and SSD in terms of detection speed,and the YOLO model also achieves good results in terms of average detection accuracy.In view of the problem that many vehicles and vehicles are easy to be detected by mistake when they are blocked in the actual road,an improved real-time detection algorithm for vehicles on the road ahead is designed.The improved YOLO model includes darknet-53 network and depth separable network.In order to improve the detection performance,the anchor box parameters are optimized,and in order to improve the detection accuracy,the anchor in the self-made data set is designed box does dimensional clustering analysis.In order to speed up the operation of the algorithm,the learning rate is constantly adjusted when the weight is optimized by gradient descent method.The experimental results show that the average accuracy(mAP)of the real-time vehicle detection algorithm based on the improved YOLO model is 1.17 higher than that of the YOLO v3 model,and the detection speed is 5 FPS higher,and the improved YOLO model can accurately detect multi-target vehicles and vehicles with occlusion.