Recognizing Context for Human-Robot Interaction

Robots and intelligent agents are becoming more important everyday. Human Robot Interaction (HRI) plays an important role in developing algorithms to utilize intelligent agents for our help. One such example is the Roboceptionist project [1] developed at Carnegie Mellon University. The Roboceptionist is a robot receptionist that does a receptionist job in a socially appropriate manner. Such socially-aware robots need the ability to detect humans in the surrounding environment, so that the behavior could be adjusted accordingly. In this paper, we describe a general object detection system that is tested on people detection tasks, face detection as well as general object detection. The system is based on the Bag of Features (BoF) approach. In the BoF approach, images are encoded in the form of normalized histograms of feature occurrences extracted using a dictionary of visual words. The count of the bucket k in the histogram corresponds to the number of features found in cluster k in the dictionary. The dictionary is built using a clustering algorithm to group similar image features into visual words. Among the experiments done in this paper, we see that using a dictionary of a large size (10K) to generate sparse histograms, has higher discriminative potential than smaller dictionary sizes (e.g 500). We also evaluate SIFT (Scale Invariant Feature Transform) features detected by the Difference of Gaussian (DoG) detector against features detected by the Maximally Stable Extremal Regions (MSER) regions on people detection. We show that MSER regions detect more stable features than the DoG detector, especially in low resolution images. Finally, we mention future work possibilities in terms of real time implementation from video frames and increasing the system’s robustness using different detectors or combinations of detectors. ∗Best viewed in color