Optimization and Improvement of Manhole Cover Status Detection Model for Mobile Road Maintenance Vehicle Based on YOLOv8

In urban road safety, missing or damaged manhole covers pose serious safety hazards to traffic, while traditional manual detection methods are inefficient and prone to miss critical issues. To address this problem, this study proposes an optimized manhole cover state detection model based on YOLOv8 and lightweighted by pruning and lightweighting technique, with the goal of deploying it on resource-constrained embedded platforms (e.g., mobile road maintenance vehicles). However, experimental results show that yolo lightweight pruning on the embedded side brings about a significant degradation in target detection accuracy (especially under high IoU evaluation). To compensate for this problem, this study adds multiple attention mechanisms: replacing DCNV2 in the DyHead module with the more expressive DCNV3 to increase the deformable feature learning capability, introducing MBConv and EffectiveSE modules in EfficientNet to optimize the C2f structure so as to improve the feature expression capability and computational efficiency, and adopting the At the same time, the lightweight Ghost_HGNetV2 is used as the backbone network to enhance the model's ability to extract and represent key features. And the size of the model is further reduced by branch reduction and lightweighting techniques, while its accuracy is improved. In order to verify the effectiveness of the proposed improvement scheme, this paper constructs a customized manhole cover detection dataset covering a wide range of environmental conditions also contains manhole cover images under different environmental conditions for evaluation. The experimental results show that with the addition of the attention mechanism, the model effectively compensates for the loss of accuracy due to pruning while maintaining a high inference speed, and achieves a good balance between inference speed and detection accuracy, which is proved to be effective in experiments. Compared with the original model, the improved model achieves significant improvement in the average accuracy (mAP) metrics of several categories.