Trajectory Prediction in Autonomous Driving with a Lane Heading Auxiliary Loss

Predicting a vehicle's trajectory is an essential ability for autonomous\nvehicles navigating through complex urban traffic scenes. Bird's-eye-view\nroadmap information provides valuable information for making trajectory\npredictions, and while state-of-the-art models extract this information via\nimage convolution, auxiliary loss functions can augment patterns inferred from\ndeep learning by further encoding common knowledge of social and legal driving\nbehaviors. Since human driving behavior is inherently multimodal, models which\nallow for multimodal output tend to outperform single-prediction models on\nstandard metrics. We propose a loss function which enhances such models by\nenforcing expected driving rules on all predicted modes. Our contribution to\ntrajectory prediction is twofold; we propose a new metric which addresses\nfailure cases of the off-road rate metric by penalizing trajectories that\noppose the ascribed heading (flow direction) of a driving lane, and we show\nthis metric to be differentiable and therefore suitable as an auxiliary loss\nfunction. We then use this auxiliary loss to extend the the standard multiple\ntrajectory prediction (MTP) and MultiPath models, achieving improved results on\nthe nuScenes prediction benchmark by predicting trajectories which better\nconform to the lane-following rules of the road.\n