The visual detection of robots is a difficult but relevant problem in several robotic applications. In this article, a framework for the robust and fast visual detection of legged robots is proposed. This framework uses cascades of nested classifiers, the Adaboost boosting algorithm, and domain-partitioning-based weak classifiers. Using the proposed framework, frontal, profile, and back detectors for AIBO robots (model ERS7), as well as detectors for humanoid robots, are built. The detection rate (DR) of the obtained systems is quite high: 90% with an average of 0.1 false positives (FPs) per image, when the final detections are filtered out using context information (horizon line). In addition, a robot referee that uses these detectors to track players during a soccer game is described. Experiment results showed that the referee achieves very high robot DRs (98.7% DR with ~ 1 false detection every 16 images) and fast processing speed.
