The detection of road cracks and potholes is an important part of road safety inspection. To address the problems of leakage and error detection in real-time road inspection, an improved YOLOv7 algorithm model was proposed. The cracks were first divided into longitudinal, transverse and alligator cracks, and then the deformable convolutional networks (DCN) was used to replace the convolution in the original YOLOv7 feature extraction network, so that the shape features of cracks with large differences in shape and irregularity were completely extracted and the accuracy of cracks was improved. The crater targets in the acquired images were generally small and not easily detected. The detection accuracy of potholes was improved by first modeling the bounding box as a Gaussian distribution, and then using a new metric based on Wasserstein distance (NWD) for small target detection evaluation method. The experimental results show that the improved algorithm improves the detection accuracy by 4.1% compared with the original YOLOv7 algorithm, while the detection speed increases by 17%, showing a better detection effect.