Dynamic 3D perception, mastery of various situations and the explainability of algorithms remain major obstacles which stand in the face of the full and complete autonomy of vehicles. Artificial intelligence laboratories are mobilizing.
The champions of the autonomous car may have lowered their ambitions, there is no question of abandoning research to achieve maximum autonomy . The main obstacles are known and leads are emerging. Starting with dynamic 3D perception, still poorly mastered by existing autonomous driving systems. “Today, we have no problem detecting, identifying and tracking the elements of a scene”, assures Jack Weast, vice president of Mobileye , an Israeli autonomous vehicle company acquired in 2017 by Intel. But this performance is only valid for a two-dimensional analysis. Until now, deep learning researchers have almost exclusively trained their visual recognition neural networks on the images provided by on-board cameras. Designed to finely detect the elements of a scene, these struggle when it comes to estimating depth. If other sensors, lidars, excellent in the exercise, “very few neural networks work on the 3D points” sent by them, recalls Michel Dhome, CNRS research director at the Pascal Institute of Clermont Auvergne University. And for good reason: they are fishing for the finesse of detection.
Bet on the fusion of multisensor data
To arrive at a dynamic three-dimensional perception – crucial to differentiate for example a real human being from a photo on a advertising panel – researchers are banking on multisensor data fusion. This consists in finding methods to combine the information coming from different sensors by using the same frame of reference, on which we will then train the neural network. If the approaches are numerous, the Vidar, a system presented by Waymo last June at the Computer vision and pattern recognition (CVPR) conference, could well “inspire other manufacturers”, according to Michel Dhome. Developed with the artificial intelligence (AI) laboratory Google Brain, it uses parallax, the effect of changing position on the observation of an object, and a mixture of data from cameras and lidars to better model the scene in a three-dimensional and dynamic way.
A welcome step forward for the future of autonomous driving, but not sufficient. While it is crucial to know exactly how far away a pedestrian stopped on the sidewalk is, it is also essential to know whether he is about to cross or whether he is waiting for a taxi. In short, to anticipate its movements. The major contractors are therefore working on methods to predict the behavior of road users. But a handful of start-ups could well beat them by combining machine learning and behavioral science. This is the case of Perceptive Automata, a young American startup founded in 2016 by a trio of researchers in neuroscience and machine learning. “Our model is trained to predict the distribution of human responses to a given situation, with physical and behavioral data,” summarizes Sam Anthony, co-founder and CTO. Installed on the vehicle’s electrical control unit, the Perceptive Automata solution receives the results of the detection system, then constructs a “virtual map” in real time where each element is surrounded by a “risk zone”, more or less smaller according to the information on the element, which changes according to its evolution in space and that of the other elements.
Make the bet of the adaptation of domain
Another obstacle on the way to complete autonomy: the creation of all-terrain systems, that is to say that will know how to drive in all conditions . A difficult task as the neural networks behind the sensors and the steering wheel were trained on a limited data set. If all the situations encountered during training are in dry terrain and in the sun, the car is unlikely to be able to drive in snowy weather. Out of the question, however, to try to multiply the data to cover all possible situations with supervised training.