Abstract: In order to improve the accuracy and reduce the model complexity of road depression detection, a lightweight target detection algorithm PSA-YOLO is proposed.Firstly, the AKConv module is introduced in the feature extraction stage, which optimizes the multi-scale feature representation by dynamically adjusting the convolution kernel and enhances the correlation of the spatial dimensions of the feature maps by combining with the ADown downsampling module.Then, in the feature fusion stage, the design of the StarECA-C2f module with adaptive channel weight assignment mechanism to effectively enhance the capture ability of fine features, and in the detection stage, a lightweight detection head LSCH is designed to reduce the computational burden by utilizing shared convolution and learnable scale layers, and combined with the WIoUv3 multi-directional weight adjustment mechanism to effectively reduce the loss of accuracy. Finally, redundant network channels are removed by Slim pruning method to further simplify the model structure and reduce the computational and storage overhead. The experimental results show that compared with YOLOv8n, PSA-YOLO reduces the number of parameters and computation by 57% and 50%, while mAP50 and mAP50-95 are improved by 2.0% and 1.9%, respectively; when tested on the GRDDC2020 dataset, PSA-YOLO's mAP50 and mAP50-95 are improved by 0.6% and 0.3%, showing the good performance and adaptability of the model on different datasets.