Predicting pedestrian trajectories plays a critical role within the context of Vehicle-to-Everything (V2X) communication technology. The main value lies in the enhancement of traffic safety and efficiency to ensure Smart Mobility. By accurately predicting future paths, vehicle can anticipate potential collisions and decrease the risk of accidents and increase pedestrian awareness. However, the prediction of pedestrian trajectories is multifaceted, caused by the complexity and unpredictability of human behaviour. Various factors contribute to the path a pedestrian follows, making it a challenging task to create precise prediction. This paper evaluates the suitability of the Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) methodology for predicting multi-modal trajectories using the Stanford Drone Dataset (SDD). The environments range from infrastructure that strongly guides pedestrian behavior up to unstructured environments. The algorithm was implemented for the paper, along with the appropriate preprocessing steps. The findings indicate that the trained model is capable of producing multi-modal predictions and forecasting short-term trajectories of pedestrians. However, the accuracy of the model’s predictions decreases in scenarios where pedestrians have greater freedom in their path choices. Based on these results the usability of MEDIRL for pedestrian path prediction has been evaluated.
Is Maximum Entropy Deep Inverse Reinforcement Learning Suitable for Pedestrian Path Prediction?
Lect. Notes in Networks, Syst.
IEEE International Conference on Cognitive Mobility ; 2024 ; Budapest, Hungary October 07, 2024 - October 08, 2024
2025-02-25
15 pages
Article/Chapter (Book)
Electronic Resource
English