A hybrid YOLOv10—Faster R-CNN framework for mobility-aid detection and traffic optimization in disability-inclusive smart cities

Efficient transportation for individuals with mobility disabilities in smart cities remains a critical challenge: high-speed detectors such as YOLO sacrifice precision under occlusion or poor lighting. Accurate models like Faster R-CNN incur latencies exceeding 100 ms per frame and lack integrated routing for disabled users. To address these shortcomings, this study proposes a hybrid YOLOv10–Faster R-CNN framework that sequentially applies You Only Look Once (YOLOv10) (operating at 45 fps) for initial mobility-aid localization and Faster R-CNN for bounding-box refinement, with a confidence-weighted fusion module to suppress false positives without compromising recall. By augmenting this dual-stage detection pipeline with an ensemble voting classifier that predicts traffic severity from refined detections and live intelligent transportation systems (ITSs) density metrics, the proposed system delivers the first end-to-end solution for real-time, accessibility-aware route planning tailored to wheelchair and crutch users—a capability previously unaddressed by standalone object-detection or traffic-management methods. We validate our approach on three complementary datasets – real-time urban traffic feeds, a diverse mobility-aid image corpus (wheelchairs, crutches), and a wheelchair-specific subdataset – and evaluate performance through mean average precision (mAP), recall, inference latency, traffic-prediction accuracy, and disabled-user travel-time reduction. The hybrid model achieves 99.4% mAP for general mobility aids and 98.9% mAP for wheelchairs, attains 100% recall (a 23.46% increase in true-positive detections over standalone baselines), and maintains an end-to-end latency of 22 ms per frame ( ≈ 45 f p s ). Traffic severity is predicted with 98.2% accuracy, and the optimized routing engine reduces disabled-user travel time by 17.3% under peak congestion compared to standard shortest-path methods. In comparative experiments, our framework outperforms YOLOv10 (mAP improvement of 2.1%) and Faster R-CNN (latency reduction of 78 ms per frame), establishing a new benchmark for inclusive, real-time traffic management in disability-inclusive smart cities. • A hybrid ITS framework integrating mobility-aid detection and traffic prediction. • Sequential dual-stage pipeline: YOLOv10 for speed and Faster R-CNN for precision. • Confidence-weighted fusion reduces false positives under occlusion and noise. • An ensemble classifier predicts congestion to enable accessibility-aware routing. • Empirical results: 100 % recall, 23.46% TP increase, 22 ms/frame latency.