Lightweight Pothole Detection Algorithm for Intelligent Chassis Control

Real-time detection of road surface potholes is crucial for implementing intelligent chassis control based on road preview, which significantly enhances vehicle driving comfort. To address the challenges of high detection difficulty and the complexity of algorithm computations associated with pothole detection, this paper introduces a lightweight pothole detection algorithm specifically developed to enhance intelligent chassis control. This algorithm is an enhancement based on YOLOv8n. Firstly, certain C2f blocks within the backbone structure are substituted with C2f Faster_ Ema blocks to reduce the weight of the convolutional blocks, thereby improving real-time detection efficiency while also cutting down the number of parameters and overall computational load. Additionally, the EMA attention mechanism is incorporated to ensure detection accuracy. Secondly, in the Neck layer, we replace conventional upsampling blocks with CARAFE sampling blocks, utilizing content-aware feature reorganization for upsampling instead of traditional nearest-neighbor interpolation methods. This substitution enhances the flexibility of the sampling kernel's position and scale, allowing for the capture of more contextual information, which in turn improves the detection performance for pothole objects. Experimental findings indicate that the proposed algorithm improves mAP@50 by 2.1% and mAP@50:95 by 1.3% on a self-assembled hazardous pothole dataset relative to baseline techniques. Moreover, it reduces the parameter count by 0.33 million and decreases the computational load by 1.9 FLOPs, thus enhancing detection accuracy while maintaining a lightweight design.