Using time-to-collision in the loss function of deep learning algorithm to improve pedestrian trajectory predictions

Raphael Korbmacher (1, 2) , Huu-Tu Dang (3, 4, 5) , Antoine Tordeux (6, 1)

1 University of Wuppertal
2 School for Mechanical Engineering and Safety Engineering
3 IRIT-SMAC - Systèmes Multi-Agents Coopératifs
4 UT Capitole - Université Toulouse Capitole
5 IRIT - Institut de recherche en informatique de Toulouse
6 JSC - Jülich Supercomputing Centre

Due to numerous real world applications, pedestrian trajectory prediction has become a hot topic in the last decades. Traditionally, researchers use physics-based models to simulate and predict the behavior of pedestrians, but recently the data-based approach has gained a lot of attention [1]. In this approach, algorithms, mostly neural networks, are trained on trajectory data to predict future pedestrian behavior. The first parts of the trajectories are used as input, often over 3.6 seconds, to predict the future trajectories over the next few seconds. The evaluation is based on Euclidean distance error metrics between the predicted and real trajectories. This approach performs well in low-density situations, where just few pedestrians are involved with long range interactions. A well-known data-based algorithm is the Social LSTM by Alahi et al. [2], which shows to outperform physics-based models in such situations. For high-density situations (2-8 pedestrians/m2), commonly referred to as crowds, the data-based approach has yet to demonstrate superior performance. We argue that this may be due in part to the limitations of the Euclidean distance metric. In high-density situations, the behavior of pedestrians is mostly driven by the avoidance of touching or colliding with other pedestrian. A more appropriate way for evaluating performance in these situations would be to take collision avoidance of the predictions into account. Kothari et al. [3] introduce a binary collision metric, however, as depicted in Figure 1 (a), it is challenging to accurately define collisions in high-density situations as touches and small overlaps are prevalent. Therefore, we propose a continuous collision error metric based on the calculation of the time-to-collision (TTC), see Figure 2 (b) [4]. By jointly training the algorithms to minimize both the distance metric and the TTC, the predicted trajectories can improve in terms of collision avoidance without altering performance in terms of distance accuracy

Mots clés: pedestrian modeling; High-density crowds; time-to-collision; Deep learning models

Domaines: Informatique [cs]; Modélisation et simulation

Raphael Korbmacher, Huu-Tu Dang, Antoine Tordeux. Using time-to-collision in the loss function of deep learning algorithm to improve pedestrian trajectory predictions. 11th International Conference on Pedestrian and Evacuation Dynamics (PED 2023), Jun 2023, Eindhoven, Netherlands. . ⟨hal-04140292⟩