Occlusion Detection in a Driving Scene By Registering Forward-looking Camera Images to a Rendered Map

Open Access
Author:
Leary, Robert D
Area of Honors:
Mechanical Engineering
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
  • Sean N Brennan, Thesis Supervisor
  • Henry Joseph Sommer Iii, Honors Advisor
Keywords:
  • map database
  • vehicle
  • occlusions
  • image registration
  • camera calibration
Abstract:
This thesis proposes a map-based vehicle occlusion detection algorithm by utilizing a vehicle's position on the road, maps of road features, and a forward-looking camera to find obstacles on the roadway. The challenge is that the mapped features must be rendered from the map database, registered to the image from the front of the vehicle, and then processed to determine occlusions. The use of a map simplifies the detection process by providing a known feature set of the road scene, where other algorithms must determine these features in real-time. The advantages of this approach include: confirmation of the vehicle's position relative to the road edge, updating of map information, rendering of map information through occluded scenes (to allow "see through" vehicles), optimization of image-processing thresholds as a function of map/position/lighting information, and provide knowledge of roads during harsh, low-visibility weather. This algorithm was successful in detecting occlusions of various road scenes, based on rendered maps of the environment.