SOA-YOLO: An Attention-Augmented, Occlusion-Aware YOLOv8 Framework for Real-Time Small-Object Detection with a Reproducible Edge-Deployment Benchmark

Real-time object detection is a core requirement for applications such as autonomous driving, video surveillance, and robotics, yet detecting small and heavily occluded objects remains an open challenge for single-stage detectors. This paper proposes SOA-YOLO, a modification of YOLOv8m that combines four components like an additional high-resolution (stride-4, P2) detection head to preserve fine-grained spatial detail for small objects, a Convolutional Block Attention Module (CBAM) inserted into the neck to refine channel- and spatial-level feature responses, a Wise-IoU (WIoU) v3 regression loss that dynamically down-weights both low-quality and already well-fit anchor predictions so that gradient updates concentrate on medium-quality anchors, and Soft Non-Maximum Suppression (Soft-NMS) at inference to reduce missed detections in crowded, occluded scenes. We evaluate SOA-YOLO on the MS COCO 2017 benchmark using a component-wise ablation study, per-object-size AP (APS/APM/APL), and a hardware-specified deployment benchmark. In a 30-epoch training run on an NVIDIA T4 GPU, SOA-YOLO improves AP[0.5:0.95] from 44.1% to 47.5% (+3.4 points) and APS from 24.6% to 29.0% (+4.4 points) over the YOLOv8m baseline evaluated under the same protocol, while sustaining 50 FPS on an NVIDIA T4 GPU (PyTorch, FP32, batch 1 versus 56 FPS for the unmodified baseline. These results indicate that targeted architectural and loss-level modifications, evaluated under a rigorous ablation protocol, can meaningfully improve small-object detection with a modest, well-characterized real-time throughput cost.