If many events continue to be canceled or postponed, the 2020 edition of the Computer Vision and Pattern Recognition (CVPR) conference, the great raid of computer vision organized each year by the IEEE, has indeed taken place – virtual – June 14-19. The 7,000 visitors could thus see that this edition gave pride of place to innovations linked to autonomous vehicles. Here are the three most striking.
Tesla details its anti-lidar approach, mainly based on machine learning
Semi-autonomous vehicle enthusiasts surely already know Tesla’s iconoclastic stance on the sensors embedded in its cars – it is the only major player in the sector to do without lidars.
What is perhaps less known is that Elon Musk’s business is also distinguished by its approach to artificial intelligence. This is what Andrej Karpathy, Director of AI at Tesla wanted to demonstrate during a workshop on the autonomous vehicle at CVPR 2020. Starting from two very similar scenes showing, on one side a Tesla vehicle and d ‘Another Waymo both taking a turn without the help of a human driver, the latter explained that the systems that allow this maneuver are actually very different for the two competitors.
Waymo – and others – rely on high-definition navigation maps with all of the geolocation data (GPS coordinates, location of signs, ground signaling, etc.), previously created using lidars. This allows the vehicle to anticipate each intersection and each change in the road environment. ” We do not make these assumptions. For us, every intersection we approach, we see for the first time “Underlines Andrej Karpathy. The images captured by the 8 cameras on board the Tesla are processed by the firm’s machine learning algorithms, which learn continuously thanks to the million vehicles sold, a figure exceeded in March 2020.
If he admits that this approach poses 100% insolvent security issues today, it is the only one that can be deployed on millions of vehicles. ” While building these lidar maps on the scale with which we operate with the sensors it requires would be extremely expensive. And you can’t just build them, you have to maintain them! “
This image recognition function is provided by the Navigation function of the Tesla Autopilot. The conference was also, for Andrej Karpathy, the opportunity to detail other functions, such as the Smart Summon, which allows the vehicle to leave its parking space to come and stand at the driver’s feet – and whose opposite function should arrive before the end of the year.
Argo AI with Carnegie Mellon to improve detection through vacuum modeling
There is no doubt that detection is one of the most critical issues for the democratization of individual vehicles with a high level of autonomy. In this area, researchers from the Robotics Institute at Carnegie Mellon University (CMU) unveiled an innovation at CVPR 2020 that could be talked about in the months or years to come, resulting from a project research supported by the start-up Argo AI.
The idea is to give a vehicle the ability to detect – and recognize – not only what it sees thanks to its sensors, but also what it does not see, hidden behind the first obstacles encountered by the lidar. Generally the recognition systems installed on semi-autonomous cars use 3D data from the lidar to represent objects as a point cloud and then try to match these point clouds to a library of 3D representations of objects. Problem, according to Peiyun Hu, who directs the research work: the 3D data of the lidar is not really in 3D – the sensor cannot see the occluded parts of an object, and current algorithms cannot take into account this second wave of obstacles, which could arise at any time.
The recognition system developed by CMU researchers is inspired by the mapping software used by Waymo, Argo AI and others – but not Telsa, therefore – by distinguishing the space occupied by an object, a human being or other, and l ’empty space “. This allows it to provide more precision in detection. ” When tested against the standard approach, the CMU method improved detection by 10.7% for cars, 5.3% for pedestrians, 7.4% for trucks, 18.4% for bus and 16.7% for trailers “Notes Carnegie Mellon University.
Waymo unveils ViDAR, a new perception system adapted to movement
The Alphabet subsidiary, Google’s parent company, took advantage of CVPR 2020 to make public and open source a high definition database consisting of 1150 scenes of 20 seconds each, well synchronized and calibrated, with data lidar and high resolution camera data, captured in a variety of urban and suburban geographic areas. The dataset contains approximately 12 million lidar case annotations and approximately 12 million camera case annotations, resulting in approximately 113,000 tracks of lidar objects and approximately 250,000 tracks of camera images. .
Above all, Drago Anguelov, director of research at Waymo, presented ViDAR, a new detection system based on cameras and distance that models the geometry, semantics and dynamics of the scene.
A collaboration between Waymo and one of Google’s many AI labs, Google Brain, ViDAR infers the structure of an object from movement. He learns 3D geometry from sequences of images – captured by cameras mounted on cars – by exploiting the parallax of movement, a change of position caused by movement. Using a few images and lidar data, ViDAR can predict future camera views and depth data.