Motion Prediction in Crowded Spaces Based on Multi-Sensor Fusion and Reinforcement Learning

Abstract

In this study, the topic of motion prediction in congested environments is addressed using multi-sensor fusion and reinforcement learning. To guarantee the reliable connection of a high-frequency LiDAR, a high-resolution RGB camera, and an Inertial Measurement Unit for real-time observation of intricate crowd behavior, a high-level system architecture has been constructed. Attention-based adaptive aggregation is the first kind of feature fusion, while a crowd-considering reinforcement learning module is the second. The suggested approach outperformed the top-performing benchmark algorithms by 19% to 24%, with an average displacement error of 17.3 cm in low-density areas and 25.4 cm in high-density areas based on the evaluation findings of the standard and real-world urban datasets. Additionally, the model has good robustness; the final displacement error in unseen, obstructed settings is still less than 27.1 cm, and the forecast accuracy decreases by no more than 12% in a noisy environment or after a sensor dropout. The technique can be reliably applied in the fields of intelligent transportation and robot navigation in unpredictable situations, according to the aforementioned trials.