HRC-YOLOv8: An occlusion and small object detection for human-robot collaboration scenarios based on YOLOv8

Abstract

In human-robot collaboration (HRC) industrial scenarios, object detection is challenged by frequent occlusions caused by human bodies or robotic manipulators, as well as by the presence of small-scale components that occupy only a limited number of pixels. To address these issues, an improved YOLOv8-based detector, termed HRC-YOLOv8, is proposed. A generalized feature pyramid network (GFPN) is first introduced by incorporating a high-resolution feature layer and multi-scale cross-layer connections, thereby enhancing shallow-deep feature interaction and improving robustness to partial occlusion. An efficient multi-scale attention (EMA) mechanism is then employed to dynamically reweight channel and spatial responses, strengthening fine-grained feature representation for small objects. In addition, an adaptive IoU (AIoU) loss is designed to optimize bounding box regression, effectively mitigating anchor-box expansion and accelerating convergence under low-overlap conditions. Experiments on the VisDrone2019 dataset and the self-constructed dataset show that HRC-YOLOv8 outperforms existing methods, validating its effectiveness and superiority in HRC scenarios.