A Prediction Method for Pedestrian Trajectory in Autonomous Driving Scenarios Based on LiDAR

To enhance pedestrian trajectory prediction accuracy in autonomous driving scenarios, this study proposes an improved adaptive interactive multiple model unscented Kalman filter (AIMM-UKF) method based on Li-DAR point cloud processing. Raw point clouds are processed using streaming techniques involving voxel grid downsampling and ensity-based spatial clustering of applications with noise (DBSCAN) clustering to extract the centroid of the minimum bounding box of pedestrians as observation input, which can enhance data reliability and real-time performance. A three-layer adaptive mechanism is introduced into the traditional interactive multiple model unscented Kalman filter (IMM-UKF) framework: time-varying transition probabilities adjusted based on likelihood functions, a weight correction factor to reinforce superior models and suppress mismatched ones, and adaptive observation noise covariance adjusted according to point cloud density to handle sparse long-range data. Experimental validation uses a VLP-32 LiDAR in a campus test scenario for pedestrian trajectories within 5 to 20 m. It shows that the proposed method reduces overall prediction error by 23.02%, and peak error during sharp turns by 29.76% compared to conventional IMM-UKF. Errors are consistently reduced by over 21% across the tested range, with the error at 20 m decreasing from 27.15 to 21.26 cm, demonstrating long-distance adaptability. Compared to state-of-the-art generative models (HSTGA, MSWTE-GNN, MPIFN), the proposed method achieves an average displacement error (ADE) of 19.3 cm, 7.21% lower than the best-performing MPIFN, while requiring only 62 ms per frame, meeting the real-time requirements of autonomous driving systems. This method enables high-accuracy, low-latency pedestrian trajectory prediction in structured environments through the synergistic optimization of point cloud streaming and adaptive multi-model filtering.