DenseXFormer: An Effective Occluded Human Instance Segmentation Network based on Density Map for Nursing Robot

Human instance segmentation in occlusion scenarios remains a challenging task, especially in nursing scenarios, which hinders the development of nursing robots. Existing approaches are unable to focus the network’s attention on the occluded areas, which leads to unsatisfactory results. To address this issue, this paper proposes a novel and effective network based on density map in the instance segmentation task. Density map-based neural networks perform well in cases where human bodies occlude each other and can be trained without additional annotation information. Firstly, a density map generator (DMG) is introduced to generate accurate density information from feature maps computed by the backbone. Secondly, using density map enhances features in the density fusion module (DFM), which focuses the network on high-density areas as well as occluded areas. Additionally, to remedy the lack of occlusion-based dataset of nursing instance segmentation, a new dataset NSR-dataset is proposed. A large amount experiments on the public datasets (NSR and COCO-PersonOcc) show that the proposed method can be a powerful instrument for human instance segmentation. The improvements of efficiency with respect to accuracy are both prominent. The dataset can be got at https://github.com/Monkey0806/NSR-dataset.

I. INTRODUCTION

Population aging has become one of the significant challenges in contemporary society. This has directly led to a shortage of nursing staff. Nursing robots are widely regarded as an effective solution to the problem and expected to generate huge social benefits. As shown in Fig. 1, human instance segmentation is a core issue among the nursing tasks, which is directly related to the safety of the care receiver. However, instance segmentation task is still a challenging task, especially in occlusion conditions.