Multi-sensor long-range object detection & classification under challenging perceptual conditions

Autonomous vehicles must be constantly aware of all aspects of the driving environment, and so are typically designed with both omni-directional and long-range forward sensor footprints. The ability to accurately detect, track and predict the motion of distant vehicles and pedestrians along the driving route remains a significant challenge, for today’s state of the art perception methods, however, despite ever-more complex network designs and ever-better sensor configurations. The inherent need in long-range detection for high resolution appearance and depth is difficult to achieve with affordable hardware, leads to real-time performance challenges, and to significant drops in detection accuracy with range to objects due to heavy occlusion and viewpoint limitations.
In this project, we will attempt to make a major breakthrough in long-range 3D object detection and tracking for autonomous driving. Our strategy will involve three components, two led by University of Toronto researchers, and one by Gatik. First, we will develop novel 3D object detection methods operating over depth ranges of over 200 m, that combine vision and lidar data to produce accurate bounding box positions. Second.

Faculty Supervisor:

Steven Waslander

Student:

Cody Reading;John Willes

Partner:

Gatik Inc

Discipline:

Aerospace studies

Sector:

Professional, scientific and technical services

University:

University of Toronto