In a response to Tesla’s Autopilot Software Director Ashok Elluswamy’s tweet linking the CVPR’22 keynote talk, CEO Elon Musk commented that it’s worth hearing about Tesla’s Autopilot progress.
Worth hearing about Tesla Autopilot software/AI progress
— Elon Musk (@elonmusk) August 21, 2022
The Conference on Computer Vision and Pattern Recognition 2022 (CVPR’22) held in New Orleans, Louisiana saw the latest advancements in the computer technology by some of the industry’s greatest. Among the hundred’s of workshops, Elluswamy gave a talk in the Workshop on Autonomous Driving, giving an insight into the latest research by the team.
Main features of the talk included obstacle detection in self driving vehicles. Done using the dense depth sensors, the visualization worked good enough at close distance. However, as the distance increased, the quality becomes inconsistent with unnecessary artifacts present.
Their solution: Occupancy Networks.
According to him, these networks acted as volumetric imaging software which uses 8 cameras to produce an image to determine occupancy. Since fundamentally everything can be dynamic after applying enough force, both static and dynamic occupancies get presented in a 3D format. The software efficiently runs at almost 100 hertz to produce uniform resolution imaging where it’s the most relevant.
Framework and examples were presented on how it’s done. Elluswamy said that their main motive for this was to avoid obstacles. Motion flow vectors were shown to describe the working of these Occupancy Networks. Through their implicit networks for collision avoidance, numerous collisions have been avoided with examples and technology presented in the talk.
Concluding his talk, Elluswamy invited researches and scientists to work together on this approach with them to make a car that never crashes.
