Posture Information Integration into Pedestrian Trajectory Prediction Considering Uncertainty

Abstract

For autonomous mobile robots, ensuring safe and efficient navigation is critically important. Especially, enhancing the accuracy of pedestrian trajectory prediction is a vital challenge for the deployment of robots in public spaces. This study sought to improve prediction accuracy by integrating pedestrian posture information into the prediction model. Utilizing a posture estimation model to gather posture data from a first-person RGB-D camera demonstrated superior performance in both Average Displacement Error (ADE) and Final Displacement Error (FDE). Additionally, we used Monte Carlo Dropout to consider uncertainty in the predictions, and we were able to significantly reduce the FDE. In the future, we plan to validate the generalization capability of the prediction model by using datasets from more complex scenarios.