Quo Vadis? Meaningful Multiple Trajectory Hypotheses Prediction in Autonomous Driving

Predicting the future behavior of traffic participants in the scene surrounding the ego vehicle is essential for the autonomous vehicle to plan a safe, comfortable, and legal route. The prediction of the other participants' future trajectories can be interpreted as a multiple-hypotheses prediction (MHP) task, due its inherent multi-modality and non-determinism. We introduce an algorithm to generate several possible future trajectories besides the trajectory that was actually traversed by the object, by utilizing the road topology encoded in the HD map data. Our introduced so-called lane-following intialization (LFI) training scheme works in two stages: First, a pretraining is performed on the generated synthetic lane-following trajectories using the evolved winner-takes-all (EWTA) training scheme, adopted to the task of trajectory prediction, second a training stage on the real data using the winner-takes-all (WTA) loss is executed. The presented LFI training scheme is evaluated and compared to training based solely on the EWTA training scheme, and to using the WTA loss training with a leakage term introduced in our previous publications, as well as to the interpretation of the problem as a single hypothesis prediction (SHP) problem, predicting only a single trajectory at a time. The evaluation is performed on the publicly available openDD dataset, providing a baseline for the task of trajectory prediction on this dataset, as well as on an internal dataset, allowing a comparison with results from our previous publication. We achieve an average Euclidean distance after 6 s as low as 1.88m on the openDD dataset, and of 1.82
$\mathrm{m}$
on the internal dataset, where previous work achieved a distance of 6.19m.