Research And Application Of Pavement Distress Detection Based On Deep Learning

With the rapid development of economy and the improvement of people’s living standards,the volume of road passenger and freight traffic has increased greatly.Once there are various diseases on the road,its use safety will reduced and road traffic safety will affected.Therefore,it is necessary to detect road diseases in real time.Traditional road damage detection is mostly carried out manually.However,with the continuous growth of road network and traffic,this time-consuming method is no longer practical.In addition,manual detection is also seriously affected by the subjectivity of decision-making.In contrast,the semi-automatic and automatic pavement detection system,which has been strictly designed and validated,can quickly and accurately detect pavement conditions and eliminate subjective effects.However,due to the complex and changeable lighting conditions,road texture and other environmental conditions,semi-automatic and automatic detection algorithms have not yet achieved great success in classification accuracy.Pavement distress detection is still a very challenging task.Based on the accuracy and real-time requirement of pavement distress detection,the pavement distress detection algorithm based on deep learning is deeply studied in this paper.The main work is as follows:(1)Improving the identification types of pavement distress detection,most of the current work is mainly on cracks or loose types(pits).From the practical application point of view,it is difficult for road managers to directly apply these studies to practical application because they do not cover all categories.Therefore,on the basis of a large number of data collection and collection,Faster R-CNN algorithm is used to identify eight types of pavement distress.(2)For the task of pavement distress detection,the network structure of Faster R-CNN algorithm is to optimized.The method of reserving small recommendation area and removing NMS in the testing stage is used to improve mAP.At the same time,on the basis of optimization,the optimal applicable parameters are compared and the best applicable parameters are obtained.(3)Analyzing the memory and computation of the model,and aiming at the real-time requirement in real-time detection,the feature network VGG-16 is compressed by combining channel pruning and low-order factor decomposition,and the detection efficiency is improved when the mAP of model detection does not decrease.(4)A pavement distress detection system is designed and implemented with multi-functional road acquisition vehicle.The software part includes the design of real-time video,local pictures,statistical query data module and interactive interface module in pavement distress detection.On this basis,the final development of the software module is completed.