Road crack detection method based on lightweight YOLOv8-LUAPD and UAV images

Abstract The rapid economic growth and heightened social activities have substantially increased road usage, leading to a higher incidence of road cracks. An efficient automated method for detecting and locating these cracks are crucial for mitigating traffic safety risks. However, existing object detection algorithms often struggle with large model parameters and high computational demands. This paper introduces a lightweight, real-time road crack detection method by enhancing YOLOv8n. Initially, a multiple-kernel convolution shuffle module is proposed to capture multi-scale contextual information during the model’s down-sampling process, thereby reducing the impact of cracks at varying scales. Subsequently, a self-calibrated local channel attention mechanism is introduced for feature selection, which not only adjusts the response weights of different channels but also incorporates local spatial information, enhancing the recognition of critical crack features. Furthermore, during the training phase of model, the bounding box regression loss function is replaced with Mpd_focaler_wiouv3, enabling the model to prioritize the regression of medium-quality bounding boxes. Finally, the improved model, YOLO-Lightweight UAV Asphalt Pavement Distress (YOLOv8-LUAPD), is evaluated against several popular detection algorithms on the UAPD. Experimental results show that YOLOv8-LUAPD outperforms mainstream algorithms in both lightweight design and detection accuracy. Notably, YOLOv8-LUAPD achieves a 6.3% improvement in mAP50 over YOLOv8n, with only 2.646 M parameters. Its inference speed on the RK3588 ARM-based edge embedded board is 33.1 frames per second, demonstrating that the proposed algorithm is well-suited for deployment on embedded devices for real-time road crack detection.