Accurate and efficient object detection is critical for road-scene monitoring systems operating in visually complex and resource-constrained environments. One-stage detectors offer high inference speed but often exhibit imprecise localization for small or low-contrast hazards, whereas two-stage detectors improve localization accuracy at the cost of high computational latency. To address this trade-off, a hybrid detection framework is introduced that reformulates the proposal stage of a two-stage detector. The method uses YOLOv5s as a learned dense proposal generator and Faster R-CNN as a region-wise refinement network. Instead of relying on the internal Region Proposal Network, high-confidence YOLOv5s detections are supplied directly to the refinement stage, and a confidence-weighted fusion strategy produces spatially consistent final predictions. The framework was evaluated on a real-world road-scene dataset using standard object-detection metrics, including mean Average Precision at IoU 0.50 (mAP@50), precision, recall, and inference speed. The hybrid model achieved an mAP@50 of 0.89, improving upon the YOLOv5s baseline by 4.7% points while maintaining near–real-time performance at 45 frames per second, approximately three times faster than standalone Faster R-CNN. The detector obtained a precision of 0.93 and a recall of 0.90 and showed substantial performance gains for small and visually ambiguous objects such as potholes and pedestrians. Multi-run experiments demonstrated consistent performance improvements with low variance, confirming the robustness of the approach. These results indicate that replacing traditional proposal generation with learned dense proposals enables a more favorable accuracy–efficiency balance than either detection paradigm alone. The proposed hybrid architecture therefore provides a practical perception solution for intelligent road-infrastructure monitoring and smart-city deployment.
Byenkya et al. (Tue,) studied this question.