| In recent years,with the rapid development of total road mileage and the rapid development of transportation industry,how to safely and accurately obtain road condition data,so as to provide timely road maintenance has become an important subject for all transportation departments.Among them,potholes and various types of cracks are the two most common types of pavement damage in the process of pavement detection.Timely detection of these damages and repair can ensure road safety and prevent traffic accidents.At present,the road condition data is mainly obtained through visual inspection by inspectors on the road surface,but there are unsafe factors and a lot of time is consumed.Based on the above problems,this thesis combines road damage detection with deep learning,and proposes an effective road damage recognition algorithm for different damage types.The main work completed in this thesis includes:(1)A road crack detection method based on improved YOLOv4 is proposed.Based on YOLOv4,the weighted bidirectional feature pyramid network was used at the neck of the network to carry out multi-scale feature fusion and integrate more feature information.Secondly,a CBAM dual-channel attention mechanism is introduced at the end of the feature fusion network to help the network better focus on the key target area and improve the detection accuracy.Finally,Soft-NMS algorithm is used to filter the prediction box to reduce the missed detection rate of the model.Compared with SSD,Faster-CNN,YOLOv2,YOLOv3,YOLOv4,YOLOv7-tiny and YOLOv5 s,it is proved that the optimized algorithm in this thesis has a better effect on road crack detection.(2)A road pothole detection method based on improved YOLOv5 s is proposed.The CA attention mechanism is integrated into the YOLOv5 s backbone network,which helps the model to locate and identify detection objects more accurately.Then Soft Pool is used in SPP module to improve the maximum pooling operation and retain more detailed feature information.CARAFE was used to improve the upsampling in feature fusion,and the resolution and spatial information of the feature map were adaptively enhanced to improve the accuracy of the model.Finally,Alpha-Io U is used to improve the loss function and improve the regression precision of the border.Comparative experimental results show that compared with SSD,Faster R-CNN,YOLOv3,yolov3-Tiny,yolov4-Tiny,YOLOv4 and YOLOv5 s,the improved YOLOv5 s model achieves the highest detection accuracy and has better detection effect.(3)Based on Qt graphical user interface development framework,a road damage detection system is designed.With the assistance of the detection system,visual display of the detection results is carried out,so that the staff can see the detection results intuitively,helping to quickly and accurately detect the type and location of pavement damage.The system includes parameter setting,damage type selection,input selection,result statistics and other functions.Finally,the system is packaged to improve the portability of the system. |
