Color Classification and Object Recognition for Robot Soccer Under Variable Illumination

Abstract

We have discussed an approach to object recognition inspired on the vision analysis pipeline. However, the linear organization of the pipeline is a model that propagates the decision of each filter, and therefore, it propagates mistakes. It is natural that after one pass over the pipeline, feedback from the result to one or several of the filters would improve the overall result. For example, finding a large circular orange ball may facilitate finding yellow goals and red AIBOs in the image since now we have information of where the ball is. Also, we have illustrated that not every pixel in the image most be processed by the entire pipeline and thus we can avoid processing some regions of the image. These remarks suggest two avenues for expanding our work. First, we can have some areas of the image significantly advanced on stages of a vision pipeline whose results may be input to other areas or early filters. Second, running the pipeline with parameters that emphasize speed but coarse results may enable further later executions of the pipeline on the same image with feedback information and adapted parameters. Having a very fast and robust pipeline here means that as CPU-speeds increase, we can run them very effectively. Notice that as resolution of the frames increases linearly, the number of pixels increases quadratically, similarly, as the frame rate increases linearly, the number of pixels that needs analysis increases quadratically. Executing a fast pipeline like ours will enable more reliable and robust systems under even larger illumination variations as we move into faster processors and more reliable hardware.