CMF-Net: Semantic Segmentation of 3D Color Point Clouds in Outdoor Scenes for Intelligent Wheelchairs Based on Multimodal Fusion

In recent years, there has been significant development in intelligent wheelchair technology, making advances in autonomous navigation, path planning, obstacle detection, and obstacle avoidance. Point cloud semantic segmentation, as a core technology for intelligent wheelchairs to perceive their surrounding environment, has become a hot topic of research. Due to the complexity of outdoor environments, traditional 3D point cloud semantic segmentation methods often struggle to accurately identify targets with similar shapes. This results in a negative impact on the navigation and safety performance of intelligent wheelchairs. To solve this problem, this paper proposes a novel semantic segmentation method of 3D color point clouds in the outdoor environment for intelligent wheelchairs by using the LiDAR and camera. This approach generates comprehensive six-dimensional color point cloud data, effectively capturing both the three-dimensional spatial and color information of outdoor scenes. Furthermore, this paper proposes a semantic segmentation network of 3D color point clouds based on multimodal fusion, called CMF-Net. It is designed to deeply integrate spatial and color features from this data, significantly enhancing the color point cloud semantic segmentation capabilities of intelligent wheelchairs. The proposed CMF-Net consists of three modules: the feature extraction module, the feature fusion module, and the feature resolution module. the feature extraction module extracts features from position, feature, and color information. This enhances the network’s ability to capture various types of information from color point cloud, thereby improving its performance. The feature fusion module processes and merges features from the feature extraction module, providing the point cloud with both local and global features. Its key advantage is the integration of an efficient channel attention (ECA) module, enhancing the focus on critical information for improved network accuracy. The feature resolution module calculates category probabilities for each point from the fused features, completing point cloud semantic segmentation. Our experimental results demonstrate that the proposed CMF-Net exhibits superior performance in outdoor environments, achieving a remarkable accuracy rate of 98.10%. Compared to PointNet, our method shows a 10% increase in accuracy, while it surpasses PointNet++ and DGCNN by 3.37% and 2.97% respectively. The proposed CMF-Net significantly enhances semantic segmentation accuracy for 3D colore point clouds. This improvement elevates outdoor environmental perception, thereby increasing the safety of intelligent wheelchairs.