Utilizing Artificial Intelligence Techniques for Car Damage Assessment and Cost Estimation

Un-anticipated road risks result in frequent car accidents, and the required assessment procedures are often time-consuming and subject to human bias. In a study that revolves around car damage assessment, an evaluation system that leverages Deep Learning techniques to analyze car damage and its estimated cost was developed by integrating YOLOv8n for vehicle-part detection, Detectron2 MaskR-CNN for severity assessment, graph-based reasoning, and cost estimation using local OEM, aftermarket and repair-price data. Furthermore, the system was trained on a public dataset (CarDD) and a private dataset consisting of 130 locally collected images, distributed into 120 images for fine-tuning and 10 raw images for raw testing. In correspondence to the study aim, re-labeling was performed by using the Label Studio tool under an expert’s supervision and guidance, to customize the public and private datasets based on the focus of detecting 11 main parts and 6 damage types. This procedure resulted in 2,523 relabeled images divided into 1,852 training, 475 validation, and 196 test images for the public dataset and 120 for the fine-tuned private dataset. Altogether, the outcome of the test data is as follows: fine-tuned YOLOv8n achieved 81.26% mIoU@50, while fine-tuned MaskR-CNN achieved 95.56% mIoU@50. In short, the end-to-end testing conducted on the remaining 10 raw cases resulted in successful valid estimates for seven cases, achieving a 70% success rate, while three cases failed due to the incapability to detect damage. A 6,053.96 SAR Mean Absolute Error (MAE) value between the estimated and reference costs was applied to the successful cases. The reference costs were derived from available pricing data rather than actual repair invoices. Overall, these results demonstrate the feasibility of an integrated simple car damage assessment and cost-estimation framework, while highlighting the need for larger real-world datasets, multiple expert evaluators, and validation against actual repair costs.