In an article published this week on arXiv, the electronic pre-publication site for scientific articles, Uber researchers presented MultiNet. It is described as a system that detects and predicts obstacle movements from autonomous laser data from vehicles. Researchers say that, unlike existing models, MultiNet reasons for the uncertainty of the behavior and movement of cars, pedestrians and cyclists by using a model that deduces detections and predictions, then refines them to generate potential trajectories.
Anticipating future levels of obstacles is a difficult task, but it is the key to preventing road accidents. In the context of an autonomous driving vehicle, a perception system must be able to capture a series of trajectories that other actors could take, rather than a single probable trajectory. For example, an opposing vehicle approaching an intersection could continue to drive in a straight line or turn in front of an autonomous vehicle; to ensure safety, the autonomous vehicle must reason about these possibilities and adjust its behavior accordingly.
MultiNet takes into account data from laser sensors and high definition street maps. He simultaneously learns obstacle trajectories and route uncertainties. For vehicles, it then refines them by discarding the trajectory predictions from the first stage and by taking the supposed center of the objects and positions of the obstacles, before correcting them and passing them through an algorithm to make final predictions of trajectory and uncertainty.
To test the performance of MultiNet, the researchers subjected the system to a day of simulations. A dataset containing readings from 5,500 scenarios collected by Uber autonomous vehicles in cities across North America using a rooftop laser sensor. They indicate that MultiNet has surpassed several baselines by a significant margin on the three types of obstacles (vehicles, pedestrians and cyclists) in terms of forecast accuracy. Concretely, the uncertainty of the simulation led to improvements of 9 to 13%, and allowed reasoning on the problem of noise inherent in future traffic movements.
“In the first case: it is an actor approaching an intersection and who has turned right. Where a basic system erroneously predicted that the actor would continue to move straight at the intersection, MultiNet predicted with great accuracy a very precise turn trajectory, while taking into account the possibility of behavior in a straight line, “explain the researchers.
Another hypothesis: “an actor made an unprotected left turn towards the autonomous driving vehicle, which IntentNet badly predicted. But conversely, we see that MultiNet has produced two possible modes, including a turning trajectory with great uncertainty due to the unusual shape of the intersection, “they added.
Source : arXiv