An Improved D-FINE Method for Concurrent Detection of Multiple Hazardous Behaviors in Elevator Maintenance Environments

To address the challenges of indistinct local details, difficult target boundary localization, and the concurrent occurrence of multiple hazardous behaviors in elevator shaft and car-top maintenance environments, this study proposes a high-resolution detail-enhanced hazardous behavior detection method based on D-FINE-S. First, we design a P2V2 shallow detail enhancement module that injects high-resolution P2 features into P3 after channel projection and spatial alignment, thereby supplementing local detail information without introducing an additional P2 detection head. Second, we propose a Residual-Adaptive Fine-grained Distribution Refinement (RA-FDR) mechanism, which applies bounded residual calibration to the non-uniform projection positions of the original FDR to improve the adaptability of bounding box regression to the data distribution of the target task. Overall, the final model achieves mAP@0.5 and mAP@0.5:0.95 values of 92.1% and 54.7%, respectively, representing improvements of 4.1 and 3.6 percentage points over D-FINE-S. The results indicate that the proposed method improves hazardous behavior detection and localization performance on the current dataset of elevator maintenance images.