Detection and Early Warning of Damaged Vehicles and Vehicles on Fire Based on Improved YOLOv8

In recent years, traffic accidents occur frequently, and the rapid detection of damaged vehicles is crucial for road safety and accident handling. However, existing research has not sufficiently addressed the detection of severely damaged vehicles, especially those with extreme deformation or on fire, which present unique visual challenges that conventional detection systems struggle with. To address this gap, we propose an improved YOLOv8 to detect damaged vehicles and vehicles on fire. Specifically: 1) we constructed a customed dataset focusing on severely deformed and burning vehicles to enhance model training for extreme damage scenarios; 2) we improved the YOLOv8 framework through several technical enhancements: a) combining with the C2fDiffTransformer module to optimize the feature extraction through the differential transformer to enhance the adaptability to complex scenes; b) introducing the CBAM attention mechanism to strengthen the expressive ability of the key feature regions and enhance the model's perception of the damage details; c) adopting the RepC3 structure to maintain real-time performance while enhancing feature fusion. Experimental results show that improved YOLOv8 detect damaged vehicles and vehicles on fire very well compared to current vehicle detection models. And our proposed method improves 7.1% on mAP@50 compared to the base YOLOv8 model and exhibits higher accuracy in detecting severely deformed and fire vehicles, inference time is 0.0343s, which meet real-time self-driving. The method proposed in this paper can provide effective technical support for accident detection and early warning in intelligent transportation systems, and help the development of autonomous driving and intelligent transportation in the future. The code and dataset has been released at Github.