Object Detection in Occluded Environments Based on Adaptive Optimal Label Assignment

The occluded environment leads to a low success rate of object detection, owing to label assignment errors and imbalances for occluded objects during neural network training process. To address this problem, an adaptive optimal label assignment algorithm is proposed, which uses a global optimal assignment strategy to minimize global loss by label assignment. It assigns an appropriate number of anchors to each ground truth, to improve the detection success rate of occluded objects. Furthermore, a focal angle loss function is proposed based on the angle loss between anchors and ground truth to achieve smooth convergence, and intersection over union loss weights are added to solve the problem of an uneven number of high-quality and low-quality anchors. Adaptive optimal label assignment and focal angle loss are integrated to YOLOv10-X, termed as YOLOv10-X-FAO. Experiments demonstrate that YOLOv10-X-FAO enhances performance over the baseline and other latest detection models. Furthermore, ablation experiments confirm the effectiveness of adaptive optimal label assignment and focal angle loss function in enhancing the detection performance of occluded targets for both anchor-based and anchor-free detectors.