Team AIBO

Brian Bentow - Yu-min Kim - Joshua Lewis

Introduction Set Up Progress Responsibilities
AIBO Pen Lab 0 and 1 Lab 2 Lab 4 Follow Ball Score Goal

CMURoboBits Lab 4 consists of creating a set of sample images which are then hand-classified to point out the important parts of the environment to the threshold generator. The generated thresholds are then used to inform higher level vision processes.

Process and Challenges

Taking Pictures

In order to get sample images, we needed to load the CMU-supplied Camera behavior and use the AIBO like a camera, using its tail to aim, and pressing a button on top of its head to take a picture. Unfortunately, the CMU-supplied code works only on the ERS-2xx model, so we had to hack it a little to get it to work. Namely, the button which stores the pictures that have been taken on to the AIBO's memory stick does not exist on ERS-7s. We changed this button to the front left foot pad.

Classifying Pictures

We took eight sample images of the pink AIBO ball. Our initial vision algorithms will only be concerned with the location of the ball (not the goal or the marker posts). To read these pictures from the memory stick log file, we used a CMU-supplied program called log-extractor. We couldn't get the pictures to extract for a while before we realized that we needed to use an ERS-7-specific configuration file as a command line option for the log-extractor tool. With the images extracted, we tiled them and loaded them into GIMP. By painting on a layer above the actual image, we hand classified pixels as either "ball" or "background" (pink vs black).

Threshold Generation

We then used the thresh program (which also had compilation issues, see Set-Up - Software), to generate thresholds based on our classifications. When we determined that our initial classifications wert insufficient, we took another set of eight sample images and classified those, running the threshold generator again.

Testing Thresholds

Finally, we took seven test images and checked whether ther ball was correctly classified. In all seven cases, it was, though the thresholds did classify some orange and red objects as "ball" also.

Results

Accomplishments

To Do