Conclusion and Future Work
It has been shown, that straight lines are appropriate for modeling the football field and the objects (except the ball) inside the view of an autonomous football playing robot. Additionally, it has been shown that a real-time algorithm can be implemented which produces a model that is close enough to the real world, in order to guide a robot through it.
Due to the fact that most engineers want to build small robots, two customary webcams are used to accomplish this goal. The sensors used in webcams are very similar to common small cameras. Unfortunately, they provided very noisy data, especially under bad lighting conditions which result in edges that often brake apart. In order to suppress this effect, the line detection tries to connect parts of lines that seem to correspond to each other.
The second part of my thesis was the construction of a robot head which is able to communicate with a computer via the serial interface or a LCD-display. The eyes can be rotated but have only one degree of freedom so far, even if the motor controller would be able to control two motors. The next step will be to mount the head onto a mobile robot which requires a wireless communication which has enough bandwidth to transmit at least 20 images per second which will result in a new 3D image every tenth part of a second. An alternative would be to do the calculations directly on the robot. A faster controller as well as more memory would be needed to accomplish this goal. Also a second motor controller will be needed to trigger two motors which will be responsible for the movement of the robot. The stereo head could also be mounted on a flying object. I considered to mount the cameras on a flying zeppelin. But there’s still the need of a wireless communication.
The algorithm could be improved, especially the line detection should additionally control the direction of the edge gradient in order to decide if the pixel belongs to a line or not. The edge detector and the line detection could be combined to reach a higher frame rate. Another improvement would be to track the objects and predict the future position. This would lead to a more stable system. Especially when few frames are reached, a movement vector (result of the prediction) can be used to interpolate between the positions. This could be useful for aiming or hitting the ball.
It was interesting to see how our visual system and, as shown, a program can derive 3D information from two views of the same area. Hopefully, the results in this area, especially in the appliance of 3D information, will lead to more intelligent robots.