A study on edge-based object detection in complex scenarios using a lightweight YOLO model

Abstract

Recent advances in artificial intelligence (AI) and intelligent algorithms have promoted the shift of object detection from cloud-based processing to real-time local perception on edge devices. To address the challenges of scale variation, occlusion interference, complex backgrounds and limited edge computing resources, this paper proposes an edge-based object detection method using a lightweight YOLO architecture for complex scenarios. The proposed model introduces a lightweight backbone network, an optimized feature pyramid fusion structure, shallow feature enhancement for small objects, attention-based feature refinement, occlusion-aware context enhancement and Soft-NMS post-processing. These strategies reduce computational complexity while improving detection stability under difficult visual conditions. To further verify the contribution of each optimization strategy, ablation experiments are conducted on the lightweight YOLO baseline. The results show that shallow feature fusion mainly improves small-object recall, the attention mechanism enhances feature discrimination in complex backgrounds, and Soft-NMS improves detection robustness for occluded targets. Compared with conventional YOLO models, the proposed method achieves 92.6% mAP@0.5, 89.8% Recall and 53.5 FPS on the edge platform, demonstrating a favorable balance among detection accuracy, inference speed and resource consumption. The method is suitable for real-time object detection tasks in edge-based intelligent perception systems.