OBJECT LOCALIZATION USING KINECT

Ambient intelligence is an emergent topic today and it involves scene understanding and object recognition. Because for scene understanding the position of objects is needed, a binary classification that decides if an object is present or not in the scene is not sufficient. The present paper proposes a system for localization of objects by a 3D bounding box. This is achieved by augmenting the 2D location, extracted using a sliding-window-based method, with 3D information acquired from a stereo-camera. While the sliding window is very computationally intensive, a branch-and-bound approach is used, which reduces the processing time, typically running in sub linear time, without discarding the optimality guarantee. An SVM is employed for classification purposes based on the bag-of-visual-word representation to obtain discrimination between object classes. The system is tested on real-world objects using a Microsoft Kinect sensor.