Object detection is an essential task for autonomous robots operating indynamic and changing environments. A robot should be able to detect objects inthe presence of sensor noise that can be induced by changing lightingconditions for cameras and false depth readings for range sensors, especiallyRGB-D cameras. To tackle these challenges, we propose a novel adaptive fusionapproach for object detection that learns weighting the predictions ofdifferent sensor modalities in an online manner. Our approach is based on amixture of convolutional neural network (CNN) experts and incorporates multiplemodalities including appearance, depth and motion. We test our method inextensive robot experiments, in which we detect people in a combined indoor andoutdoor scenario from RGB-D data, and we demonstrate that our method can adaptto harsh lighting changes and severe camera motion blur. Furthermore, wepresent a new RGB-D dataset for people detection in mixed in- and outdoorenvironments, recorded with a mobile robot.
展开▼
