Vehicle Trajectories from Unlabeled Data Through Iterative Plane Registration

Abstract

One of the most complex aspects of autonomous driving concerns understanding the surrounding environment. In particular, the interest falls on detecting which agents are populating it and how they are moving. The capacity to predict how these may act in the near future would allow an autonomous vehicle to safely plan its trajectory, minimizing the risks for itself and others. In this work we propose an automatic trajectory annotation method exploiting an Iterative Plane Registration algorithm based on homographies and semantic segmentations. The output of our technique is a set of holistic trajectories (past-present-future) paired with a single image context, useful to train a predictive model.

Multiple Future Prediction Leveraging Synthetic Trajectories

“…Alternative ways to gather trajectory data is to rely on less expensive existing videos lacking sensor annotations and trying to estimate vehicle motion, for instance using SLAM [14] or replacing sensor data with deep learning methods [15]. These methods however still require high quality videos captured from a moving vehicle.…”

Explaining Autonomous Driving by Learning End-to-End Visual Attention

“…Simulators are often used to develop and test such models, since machine learning algorithms can be trained without posing a hazard for others [15]. In addition to learning to transform visual stimuli into driving commands, a vehicle also needs to estimate what other agents are or will be doing [35,4,30,32].…”