Aqua Phoenix
     >>  Research >>  Sentinel  
 

Navigator
   
 
       
   

5.4.2 Priority Ordering

Commands that are more complicated in nature should be ordered higher on the priority of checking and making decisions. Left and right are extremely broad definitions and can be placed at the bottom of the hierarchy. Aside from moving straight, they are the base operators for controlling the tank.

Here is the order of command checking as follows:

  1. Is it a fork?
  2. No? Is it a left?
  3. No? Is it a right?
  4. No? Go straight.

5.4.3 Memory Heuristic

Rather than changing decisions on the fly, based on what image is given to the decision algorithm, it is more worthy to store a memory of what command should be done as the next command. Humans, when they see a road veering off to the right, do not immediately turn their vehicles and make a hard right turn. Humans often keep in mind that they need to make a turn in a particular direction, and they make the actual decision of turning when the appropriate time comes around.

The way our algorithm makes the actual change of decision - the decision to go from straight to left or from right to straight and so forth - is based on what the camera saw last. How does the decision algorithm know when to make the decision? By seeing a space that is mostly covered with background. In our experiment, the Sentinel did not get lost once the images from the camera were composed entirely of background. It simply processed its memory - which was controlled by the last image the algorithm saw. Below is an example of what occurs during a left turn. (M=memory, D=decision)

Figure 1.12: Memory Heuristic
There could be incorrect analyses saved to memory based on the images seen. For example, seeing one image could set the memory to "left" while the next image sets the memory to "right." Although there could be a vacillation between decisions, the last image seen will be the deciding factor, before the entirely black image is processed and a real decision must be made. This reinforces the fact that a road may look like it is veering left at first, but the left was very slight compared to the huge right turn coming up.

5.4.4 Thresholding

Thresholding greatly assisted in the Sentinel's driving skills. Recognizing forks became a problem when only observing a border thickness of 1 pixel. Five pixel borders worked out much better. The minimal threshold area distinguishes between adjusting the memory variable or actually making a real decision. It is not good to make the threshold equal zero, because images from a wireless web cam tend to show at least a few stray white pixels, and this will negatively influence our algorithm.