Research On Road Surface Disease Detection Method Based On Improved YOLO Network

With the continuous improvement of China’s infrastructure,the road traffic system has been developing steadily year by year,gradually forming a huge traffic network system.At the same time,as the frequency of road use increases,road maintenance issues are receiving more and more attention.Fast and accurate detection of road damage targets can help road departments or relevant road maintenance agencies to repair road surfaces in a timely manner,thus ensuring road safety.In this paper,we use part of the publicly available dataset RDD2020 and the road surface disease dataset collected by ourselves in four months for training and testing,and investigate how to improve the accuracy of road surface disease detection under complex environment.The research content is as follows:First,in order to improve the accuracy and robustness of the model after YOLO training in subsequent experiments,this paper uses multiple data enhancement techniques to enhance the road surface disease dataset,expanding the dataset to 6120 images.A combination of enhancement techniques is used,including image rotation,mirror flip,random cropping and scaling,cutout,adjusting brightness and contrast,and Gaussian blur.Compared with a single enhancement method,the combined data enhancement can generate more diverse and richer training data,improve the generalization ability and robustness of the model,and reduce the risk of overfitting.Secondly,in order to solve the problems of missed detection and insufficient detection accuracy of YOLOv3 in the detection of disease characteristics of small targets with complex shapes under complex background,this paper proposes a road surface disease detection algorithm based on SMC-PHD and YOLOv3.By combining SMC-PHD filter and YOLOv3 network,the detector uses SMC-PHD filter to estimate the state of the target,further improving the accuracy and robustness of the target detection.At the same time,the state estimation results are fed back into the object detection process,enhancing the algorithm’s overall ability to identify road surface disease targets and further improving the accuracy and robustness of road surface disease detection.Finally,to address the shortcomings of the original YOLOv3 network in terms of lightweight degree and feature extraction,this paper proposes an improved model,YOLOv3-SNCA,based on Shuffle Net V2 network structure and CA attention mechanism.This paper also uses a loss function called Focal Loss to solve the problem of ranking a large number of bounding boxes in the prediction stage.Compared with other common loss functions,Focal Loss can effectively mitigate the influence of those easily classified samples on the training of the model,and focus more on the learning of samples difficult to classify,thus improving the performance of the model.On this basis,this paper combines the SMC-PHD filter introduced in the previous paper and proposes a road surface disease detection algorithm based on SMCPHD and YOLOv3-SNCA to further improve the detection accuracy and stability of the model.It is demonstrated through experiments that the proposed SMC-PHD filter and YOLOv3-SNCA based road surface disease detection method solves the problem of low detection accuracy and slow speed of the original YOLOv3 target detection network in road disease detection tasks in complex backgrounds,improves the accuracy and real-time detection performance in complex environments,and can provide help for road maintenance work.Therefore,the research is of great research significance.