A 3D Point Cloud Human Body Parts Segmentation Using U- Net Based on Spherical Projection

Abstract

Light Detection and Ranging (LiDAR) has become crucial in human activity recognition. LiDAR reflects and receives laser beams from humans, including surrounding objects, visualizing the reflected results as a 3D Point Cloud. This often faces challenges like object shape variations and noise that can cause overlap, hindering segmentation and object recognition accuracy. The spherical projection approach is used to address these issues. This paper proposes a spherical projection for the segmentation approach, which comes as an answer. The methodology employed in the study focuses on effectively partitioning data derived from LiDAR technology, leveraging a unique spherical projection technique. This innovative approach involves the transformation of a three-dimensional Point Cloud into a two-dimensional projection, facilitating more efficient segmentation processes. Integrating U-Net, a deep learning model recognized for its prowess in image segmentation tasks, aims to enhance the accuracy and precision of delineating human body parts from LiDAR-generated data. The proposed method contributes to advancing the segmentation technique field and highlights the potential applications of U-Net in LiDAR data preprocessing. The model achieved an accuracy of 99.74% and Intersection Over Union (IoU) results of 87.92%.