Door Handle Detection for the Stanford AI Robot (STAIR)

We adapted a visual object detection framework, originally applied to face recognition, to perform fast and robust door handle detection. We detect objects in real time using a cascade of weak classifiers. The framework rejects most image subwindows early on in the cascade and a very small minority ever reaches the later nodes. Our system achieves a high detection rate (approximately 95%) while maintaining a very low number of false positives, even though door handles are intrinsically less feature-rich than faces.