next up previous
Next: Camera Calibration Up: Stereo vision based Previous: System Architecture

Stereo

Spinoza uses a trinocular stereo system for sensing. Stereo provides dense depth maps and has high angular resolution. This allows it to resolve small or narrow objects significantly better than sonar arrays. Laser range scanners obtain a high degree of accuracy but require a long time to scan an area. Stereo vision is also a passive sensing technique, which is preferable to active sensing for some applications. One of the major limitations to the use of stereo has been its high computational cost. Spinoza overcomes these limitations by employing high--speed convolvers to ensure fast computation and precision calibration to achieve depth images with little noise.

The system gives a high quality depth map that is 128x120 pixels with 20 disparities and computes in 350 ms. This very high performance combined with wide angle lenses allows the robot to see a large area very quickly so that it can map without pausing and move at an acceptable velocity.

The stereo system computes depth maps using three calibrated cameras with an algorithm similar to the multi baseline stereo [10]. On Spinoza the first pair of cameras are in a horizontal plane while the second are in the vertical plane. The trinocular cameras can achieve better results than a typical two camera stereo system because the second pair of cameras can resolve situations that are ambiguous to the first pair. For example when the horizontal camera pair looks at a scene that has only horizontal lines, the depth of the images is ambiguous, but the vertical cameras can find the depth for this scene.





Vladimir Tucakov
Wed Dec 4 11:45:59 PST 1996