Hybrid Attention-Guided YOLOv5 based Object Detection Framework with Cross-Modality Fusion and Edge Optimization

This research introduces Hybrid Attention-Guided YOLOv5 (HA-YOLOv5), a cutting-edge object detection framework that harmonizes multi-sensor intelligence with deep learning efficiency to redefine real-time perception. By integrating sparse attention-based fusion of RGB, thermal, and depth data, the system dynamically prioritizes critical features while suppressing noise, enabling robust detection in challenging conditions like low-light or occluded scenes. A CBAM-boosted backbone and reinforcement-learned gating (GFP-Neck) optimize feature flow, achieving unprecedented accuracy (92.3% mAP) without compromising speed (120 FPS). HA-YOLOv5 outperforms state-of-the-art models (Faster R-CNN, YOLOv5s, EfficientDet-D2) by 5–14% in accuracy while maintaining real-time edge deploy ability through mixed-precision quantization. Unlike conventional detectors, its self-supervised auxiliary head (SSA-Head) learns from unlabelled data, reducing annotation dependence. The system bridges the gap between precision (91.8%) and adaptability (92.7% recall), proving especially effective for autonomous driving, surveillance, and medical imaging. Beyond metrics, HA-YOLOv5 rethinks AI-assisted perception—not just detecting objects, but understanding multi-sensor contexts with human-like attention. This work demonstrates that the future of vision systems lies in hybrid architectures where attention mechanisms guide computational efficiency.