Research On Feature Extraction And Recognition Algorithm Of Pavement Potholes And Upheavals Diseases Based On 3D Structured Light Detection Technology

For the dominating damage types of asphalt pavement,potholes and upheavals are not only affect the city landscapes,but also impacted on road safety.If not repaired in time,the above-mentioned diseases are likely to further develop and cause structural damage to the pavement,shortening the life cycle of the road.For the traditional image recognition method,it is easy to be interfered by noise and other factors in the process of extracting the disease features of potholes and upheavals,thus causing more misjudgments.Three-Dimensional feature extraction imaging method has the problems of high detection cost and poor interoperability of detection data.In order to further improve the recognition accuracy and efficiency of pavement potholes and upheavals,aiming at the similar feature of texture features in two-dimensional image,3D structured light surface scanning technology is used to obtain the potholes and upheavals image containing three-dimensional information,and convolution is introduced to propose a new potholes and upheavals classification model.First and foremost,the asphalt pavements of municipal road were selected as the test section in this paper,using the laser surface scanning technology to obtain the 3Dimensional high-precision pavement data.Combined with the two-dimensional image,the RGB image with high-precision three-dimensional information is generated,and the three-dimensional disease database is established.Then,based on convolution neural network technology,the classification and recognition model CNN 1 and feature extraction model CNN 2 are established,and the model was optimized.Finally,potholes and upheavals diseases images obtained through screening of CNN 1 model were taken as input samples of the CNN 2 model.Further extract the area and elevation information of the potholes and upheavals.The results show that the CNN 1 method in the Validation set,classification accuracy rate,recall rate,precision rate and the F1 comprehensive evaluation value were 97.34%,96.34%,99.85%,and 98.12%,respectively.In the test set,all the indicators reached 97%,the average time under GPU acceleration was 10.32ms/sheet.The accuracy indexes of object detection of potholes and upheavals CNN 2 of m AP₅₀,m AP₇₅ and m AP_50-95were0.912,0613 and 0.560,respectively.The average time under GPU acceleration is1.19/sheet.CNN 1 model can significantly improve the classification and recognition accuracy of pothole and upheaval,CNN 2 model can accurately obtain the area and elevation information of the pothole and the upheavals,the combination of CNN 1 and CNN 2 effectively improves the detection efficiency and accuracy of pavement potholes and upheavals.