EG-PointNet: Semantic Segmentation for Real Point Cloud Scenes In Challenging Indoor Environments

When patients with movement disorders used a robotic arm for supported living, the target objects in complex surroundings offered various and overlapping states, making it difficult for the robotic arm to grip the target objects precisely and effectively. As a result, to solve the aforesaid difficulty, we designed a point cloud recognition method for complex scenarios. In this paper, we proposed EG-PointNet, a novel neural network that adds the EGconv module to perform semantic segmentation of indoor point cloud data. EGconv provided several distinct advantages as compared to current modules that execute in outer space or treat each point separately: it combined local neighborhood information; it could be stacked apply to learn global shape attributes; it could capture semantic properties over possibly large distances in the original embedding in multi-layer systems because of its affinity in feature space. Experimental results identified that the proposed EG-PointNet could be successfully integrated into a robotic arm vision module to help patients with movement disorders grab target objects accurately in challenging environments, proving our network’s effectiveness.