Three-dimensional (3-D) point clouds are 3-D spatial image data composed of a large number of 3-D coordinate points, which are usually detected by sensors, light detection and ranging (LiDAR), or other means. Human motion prediction and analysis investigate the spatiotemporal dynamic changes of biomechanical parameters during movement, including limb position, velocity, and acceleration, thus playing a pivotal role in multiple fields such as sports training, rehabilitation medicine, and ergonomic design. However, existing research on human motion analysis predominantly relies on physical sensors and video streams for data collection. These conventional methods often fail to generate precise 3-D models. Most outputs are limited to 2-D images or spatial trajectory data, which have significant limitations in analyzing subtle spatial pose details during dynamic motion processes compared to comprehensive 3-D models. In traditional deep learning methods, the processing of 3-D models often requires excessive computational resources.To perform human motion prediction on 3-D models without view constraints and save computational resources, thisstudyintegrated 3-D LiDAR point clouds with human motion analysis to overcome the inherent limitations of conventional 2-Ddataacquisition in achieving precise modeling. Additionally, this study adopted a novel temporal prediction algorithm frameworkspecifically designed for 3-D humanpoint cloud sequences, achieving motion prediction with significantly reduced memoryoverhead.

Firstly, an optimized convolutional neural network (CNN) architecture was adopted to perform multi-round iterative calculations on sequential 3-D human point cloud data, so as to achieve data dimensionality reduction and multi-level feature extraction. The network reconstructed traditional convolutional and pooling layers into specialized structures adapted to 3-D tensors, thereby effectively capturing the spatial topological relationships within the point cloud data. On this basis, combined with the principal component analysis for dimensionality reduction and the sparsification algorithm based on average distance, the feature dimensions were further compressed to obtain low-dimensional feature tensors with high discriminability. Subsequently, the K-nearest neighbor-based density estimation method was applied to conduct cluster analysis on the reduced-dimensionality features, aiming to identify the potential pose distribution structures within the point cloud data. Based on the clustering results, a hybrid model consisting of multiple fully connected neural networks (FCNN) and long short-term memory networks (LSTM) was constructed for temporal feature modeling and motion trend prediction. This hybrid structure gave full play to the strengths of FCNN in feature mapping and the advantages of LSTM in long-range dependency modeling. In the stage of feature prediction and pose reconstruction, the system adopted a parallel processing architecture to run multiple FCNN-LSTM modules simultaneously, with each module independently handling different subtasks, thereby significantly improving computational efficiency and resource utilization. Finally, an independent FCNN mapped the predicted features back into 3-D space to reconstruct the corresponding human pose point cloud representation.

In this study, four groups of human point cloud motion data of different populations were collected to verify the proposed method (Table 1, Fig.4), and four groups of different experiments were designed. These experiments were conducted from four different dimensions to verify the effectiveness and efficiency of the proposed method, namely different poses (Fig.5), different time frames (Fig.6), different densities of point clouds (Fig.7), and different iteration numbers (Fig.8). In addition, the proposed method was compared with other human motion prediction methods in recent years to further verify the effectiveness and efficiency of this algorithm.The overall process achieved efficient andscalable temporal human motion modeling and prediction while maintaining high accuracy.The overall process achieved efficient and scalable temporal human motion modeling and prediction while maintaining high accuracy.

The algorithm proposed in this study can quickly and accurately predict the future poses of human body during movement with low computational resource overhead. Moreover, it shows better performance on data with larger motion range and smaller degree of point cloud sparsity. In general, the algorithm in this study well balances computational efficiency and recognition accuracy, and provides better support for real-time motion prediction applications with limited resources. It has unique application value in fields such as human-machine collaboration optimization, sports training, and rehabilitation medicine.