Real Time Obstacle Detection and PredictionMembers: Minwoo Park, Robert Collins |
Project Descriptions
Independent moving object and close objects are detected from a moving pair of stereo cameras. Image sequences are taken by stereo pair of the cameras attached to the front of the vehicle. Next, the depth and the optical flow are computed from the sequences. Using analysis upon these 3D optical flows, close objects and possible collision are detected. One of the main advantages of this approach is that it does not need any spurious assumptions and it can provide with better segmentation due to rich information. The accuracy of detected change of the depth is the key component.The procedures are the below.
- Calibration
- Scene Acquisition
- Undistortion: inference of the inverse function, because it can't be evaluated in closed form
- Feature detection of strong edge value
- LK tracking over the frame and the pair image
- Computation of the absolute distance Z and inference of the world coordinate X and Y
- From the information found above, clustering of the 3D flow is performed
- Every cluster is segmented using stereo information over edge
- Mean of the 2D velocity vector is assigned for the every segmented clusters
 [stereo object detection] |
 [calibration] |
 [flow] |
 [flow segmentation] |