Learning and predicting on-road pedestrian behavior around vehicles

Pedestrian safety is critical from the standpoint of autonomous vehicles. We address the problem of reliable long term (<; 5 sec) prediction of on-road pedestrian trajectories using track histories generated with two vehicle mounted cameras. Due to the relatively unconstrained nature of pedestrian motion, model based trajectory prediction approaches become unreliable for long term prediction. Data-driven approaches offer a viable alternative, but are susceptible to bias toward dominant motion patterns in the training set. This leads to poor prediction for under-represented motion patterns, which can be disastrous from a safety perspective. In this work, we extend the Variational Gaussian Mixture model (VGMM) based probabilistic trajectory prediction framework in [1]. We sub-categorize pedestrian trajectories in an unsupervised manner based on their estimated sources and destinations, and train a separate VGMM for each sub-category. We show that the sub-category VGMMs outperform a monolithic VGMM of equivalent complexity, especially for longer prediction intervals. We further analyze the errors made by the two models and the distributions learnt by them, to demonstrate that the sub-category VGMMs better model under-represented motion patterns.