To recognize Objects we use a Single Shot Multi-Box Detector which is one of the fastest state of the art convolutional-neural-network based algorithms. It allows us to efficiently and accurately scan the Xtion video stream for occurrences of specified items. We employ an implementation from the Tensorflow Object Detection API.
We train it with a large amount of automatically generated training images, rendered via a customized blender pipeline.
Generated training images for screwdrivers and the battery object from the SpaceBot Cup (the two on the left) and a detection-result of the trained algorithm (on the right)