Learning to Estimate Multivariate Uncertainty in Deep Pedestrian Trajectory Prediction

With the advent of autonomous vehicles (AVs), it is mandatory to care for pedestrians' integrity, as they are one of the most vulnerable entities in transit. Therefore, the AVs must anticipate their actions and predict their trajectories to improve tasks such as active perception, predictive path planning, predictive control, and human-robot interaction. The literature presents deep learning methods to predict pedestrian trajectories from the perspective of an onboard camera. However, only one study modeled the uncertainties involved in the model prediction. Thus, we address the problem by proposing a method to model both aleatoric and epistemic multivariate uncertainties in deep pedestrian trajectory prediction. We are the first to model the multivariate predictive uncertainty in pedestrian trajectory prediction by incorporating mathematical conditions to ensure stability during training. Our methodology can be applied to any deterministic method with minimal adjustments and present more accurate results than the BayesianLSTM.