YOLOv7-RDD: A Lightweight Efficient Pavement Distress Detection Model

The increases in the timeliness of maintenance and repair requirements for urban road pavements have created a demand for the timely detection of pavement distress. An effective detection method of multiple types of pavement distress based on low-cost front-view video data was proposed. We constructed a target detection model YOLOv7-RDD modified You Only Look Once version 7 for Road Detection by modifying the YOLOv7 framework to reduce the complexity of the model while improving the detection precision of multiple types of pavement distress. The lightweight efficient aggregation network structure using distributed shift convolution (EDC), the improved spatial feature pyramid structure (SPPCSPD) proposed, and the similarity-based attention mechanism (SimAM) were integrated into our model. We created a dataset called CQURDD containing 12 pavement distress types for urban roads. The SPPCSPD structure has been proven to reduce the running time by nearly 54% compared to the spatial pyramid pooling structure in YOLOv7 while losing little precision. The introduced SimAM module and SIoU (Smooth Intersection Over Union) loss function have improved the model precision, and the Non-Maximum Suppression (NMS) can help reduce redundant detection boxes. A comparative study was carried out based on the CQURDD dataset, YOLOv7-RDD achieved the best precision and efficiency with mAP@0.5 reaching 89.5% and FPS reaching 145. Except for linear cracks, the AP values for the other types of distress detection are all greater than 85%. Moreover, the excellent performance on deformation type distresses of our method forms a relative advantage over professional top-view detection.