|
|
Coyote Robotics Technology Our technology approach divides the problem into these sub-areas: terrain and obstacle detection, road and lane tracking, vehicle motion estimation, and local navigation (vehicle control planning). To accomplish these functions the system employs stereo vision and single camera moving obstacle detection, multi-cue probabilistic road and lane detection, visual motion detection for determining vehicle speed and direction, probabilistic SLAM (simultaneous localization and mapping) to reduce map error and refine precision, and a discrete state engine to make local navigation decisions. Our approach emphasizes the use of cameras over other sensors such as LADAR. Camera systems can utilize different lenses (and different baselines in the case of stereo) to extend range whereas LADAR must provide a substantially higher power beam. The components of the stereo system also cost less than a comparable 3D LADAR system. The images used to generate stereo can also be used for other vision processing tasks such as lane following and moving object detection and tracking. By sharing the image between multiple tasks, all vision processing tasks can compute results for precisely the same instant in time. The vehicle sensors are positioned as follows:
The vehicle uses two fixed camera pairs that process the forward and backward looking view. In addition to these camera pairs, the rotating gimbal camera processes long range image data. Obstacle detection is accomplished by using the stereo disparity (depth) map to determine location, height and size of obstacles:
From the left and right input images (left), a disparity map is generated (middle), which is further processed to mark obstacles (right). Lane and road detection is performed by reusing one of the stereo input images. Lane detection detects white and yellow lines by using edge detection. Road detection is performed through color probability matching:
While the components of the vision system feeds sufficient info into the Local Navigation component, the GPS with supplementary INS are responsible for orienting the vehicle globally and for finding the optimal path in the RNDF. The path finding algorithm continuously plans and re-plans the path if necessary to find the fastest time to the next waypoint. The Local Navigation system makes all of the decisions regarding when and where the vehicle will drive. The system does this using state machines to represent the vehicle’s current state and following a set of rules that dictate the vehicle’s action.
|
Copyright (c) 2004-2006 Coyote Robotics, LLC
