Neural Networks for Visual Object Recognition Based on Selective Attention

Ulrich Kaufmann, Rebecca Fay, Heiner Markert, Günther Palm

Object localisation and identification is an important crucial problem for advanced mobile service robots. We implemented a neurobiologically plausible system on a robot that localises and identifies objects using a colour-based visual attention control algorithm and a hierarchical neural network for object classification. Object localisation and classification are performed in two stages. First low-resolution features are used to determine windows of attention in the robot’s camera image. Then highresolution visual features are used for object recognition. The approach is evaluated in a test scenario where a robot is located in front of a table carrying different objects. The robot has to identify and manipulate these objects. We evaluated the total object recognition performance and compared the effectiveness of different feature sets. The approach performed very well regarding object localisation and classification results in this scenario and meets real-time constraints.