Depth-based detection using haar-like features
Ruud Mattheij, Eric O. Postma, Y. van den Hurk, Pieter H.M. Spronck · Research portal (Tilburg University) · 2012
The automatic detection of objects has gained considerable attention over the last few years. Most objectdetection approaches rely on visual features that are sensitive to identity-irrelevant variations, such as changes in illumination. Being less sensitive to such variations, depth features may improve detection accuracy. Depth features can be extracted from depth images generated by commercially available depth sensors, such as Microsoft’s Kinect device. This paper describes a method for robust and accurate face detection by employing Haar-like region features on the integral image representation of depth images. Our aim is to determine to what extent region-comparison features contribute to effective face detection in depth images, compared to pixel-comparison features. To this end, we present a revision of the recently proposed detector of Shotton et al. [10]. Whereas the detector of Shotton et al. relies on pair-wise pixel comparisons in depth images, our revision compares square regions in a pair-wise fashion. In a comparative evaluation of the original and revised method, we train and evaluate both detectors on our depth images of faces (DIOF) database that is compiled at our lab. The results reveal that the use of region features instead of pixel pairs indeed improves face detection accuracy in depth images. We conclude that employing region features contributes significantly to effective face detection. Future work will address to what extent our results generalize to the detection of body parts and objects in general.