Research And Application Of Road Damage Detection Algorithm Based On Convolutional Neural Network

With the continuous improvement of the road network,the demand for road maintenance is growing,while the traditional human-vehicle patrol mode or walkingon-the-road mode is inefficient and subjective.The development of pavement maintenance urgently requires a fast,efficient and low-cost road damage detection method.At present,computer vision technology based on convolutional neural network has become the mainstream,using convolutional neural networks to learn road damage features to identify objects,which is a common means of automatic road damage detection.However,in practical applications,due to the complex background of road image,the detection accuracy of the existing algorithm is low and the application effect is poor.To solve the above problems,this thesis studies the road damage detection algorithm based on object detection and object tracking,and uses TensorRT to optimize the algorithm so that it can be transplanted to an embedded terminal to achieve realtime road damage detection,and then develops a road damage intelligent analysis system to achieve visualization of the detection results.The specific work is as follows:(1)In order to improve the detection performance of the model,on the one hand,the existing road damage dataset was optimized and expanded: on the basis of the RDD2020 dataset,real road images were collected and annotated,then expanded to the dataset,and label categories were added to optimize the annotation in the dataset to finally establish a high-quality XZ-RDD road damage dataset;On the other hand,based on the lightweight model YOLOv5 s,the RepCA-YOLO model was proposed,the CA attention mechanism was introduced,and two multi-branch C3 structures were designed by drawing on the RepVGG idea.The structural reparameterization technique was used to ensure the size and real-time detection performance of the model,without increasing the number of parameters while strengthening the model’s ability to extract road damage features.The experimental results show that the accuracy of the RepCAYOLO model reaches 70.1% on the RDD2020 dataset,with an improvement of 3.3%,and 81.0% on the XZ-RDD dataset,with an improvement of 3.7%.(2)In order to meet the road damage recognition requirements of object reidentification in practical applications,the target tracking algorithm is used to achieve object re-identification.Firstly,the road images were collected using the vehiclemounted camera and annotated to produce the road damage tracking dataset RDT.Secondly,to improve the tracking accuracy of the model,the Re ID network was improved on the basis of the DeepSort target tracking algorithm by deepening the depth of its network and increasing the output dimension,so as to improve the network’s ability to characterize the damage features for the matching of trajectories and detection boxes.The experimental results show that the tracking accuracy of the model on the RDT dataset reaches 58.9% and the precision reaches 74.5%,an improvement of 2.8%and 2.2% respectively.Finally,RepCA-YOLO is used as the object detector and the improved DeepSort as the object tracker to achieve road damage detection with object re-identification.(3)In order to realize the practical application of the project,the above model was transplanted to the embedded terminal.Firstly,the model was optimized using TensorRT technology to accelerate the model inference to ensure that the model can detect road damage in real time on a terminal with limited computational resources,and the optimized inference model was deployed to the Jetson NX development board.The experimental results show that the inference time of RepCA-YOLO model optimized by TensorRT on the development board reaches 8.6ms and the detection accuracy reaches 80.7%,and the model combined with the improved DeepSort achieves an inference time of 14.5ms and a tracking accuracy of 56.1%,meeting the requirements of real-time detection.Secondly,the in-vehicle intelligent terminal device was developed to deploy the above inference model,and the device was installed on the vehicle to obtain real-time road images through the camera and detect road damage,and obtain real-time road damage detection results and upload them to the server.(4)In order to visualize the detection results,the road damage intelligent analysis system was designed and developed.Firstly,according to the system requirements,the relevant functions of the system were designed and developed to visualize the damage report and feedback the damage status.Secondly,the intelligent terminal was deployed on the shuttle buses and patrol vehicles respectively,and the application tests were done on S253 and G104 routes to verify the application effects of the terminals and the system,and finally the automated road damage detection is realized.