In the figure below, the optical flow field computed on two successive images of our lab taken as our robot was moving forward with constant speed is shown. The flow field is also shown superimposed on the image:
Everything that is visible through the camera lens have a depth information. But we can make use of the camera position and orientation to filter out the floor, and to detect the obstacles that are visible. In the figure below, the red colored places are visible, and we can segment the part belonging to the chair easily:
By using the same optical flow field computed above using two images from our lab (also shown below again), we can find the depth information for each flow vector by combining the time to contact computation, and the robot's speed at the time the images are taken. Below is a depth map for the same flow field drawn in 3D (and shown from various directions to reveal the 3D structure) :
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |