Hand Tracking based on Hierarchical Clustering of Range Data

摘要: AbstractFast and robust hand segmentation tracking is an essential basis for gesturerecognition thus important component contact-less human-computer inter-action (HCI). Hand gesture recognition based on 2D video data has been intensivelyinvestigated. However, in practical scenarios purely intensity approaches sufferfrom uncontrollable environmental conditions like cluttered background colors.In this paper we present a real-time algorithm us-ing Time-of-Flight (ToF) range cameras data.The rangeinformation fused into one pixel value, representing its combined intensity-depthhomogeneity. The scene hierarchically clustered using GPU parallel merg-ing algorithm, allowing identification of both hands even inhomogeneousbackgrounds. After the detection, are tracked CPU. Our trackingalgorithm can cope with situation that temporarily covered by otherhand.1. IntroductionGesture-based interaction requires fast ofthe human [13]. Classical or color images. However,this kind techniques suffers from low efficiency lack robustness case clutteredscenes if applied under varying lighting conditions. Addressing application scenarios,techniques capable handling effects strongly required; frequently simplifica-tions, e.g. restricted material [9], marker- glove-based [16]are hardly applicable.One major approach to overcome problems segmenting image sequencesfor purposes use additional depth information, delivered laser rangesystems [7], stereo [11] structured light acquisition systems [12]. majordrawback all these comparably expensive sensing hardware sig-nificant space requirement, which due systematic constraints, baseline required forstereo light, mechanical setups scanners.Recently, (ToF)technology, measuring time emitted byan illumination unit travel object back detector, manufacturedas highly integrated ToF cameras. Unlike other 3D systems, arevery compact. ToF-cameras realized standard CMOS CCD technology becost efficiently manufactured [10, 20]. have successfully contextof man-machine interaction, facial [6], touch-free navigation medicalapplications [15], upper-body-gesture [8] hand-gesture [4].In paper, introduce hierarchical clusteringtechnique. Using clustering beneficial, since final number clusters thescene delivering “best” segmentations depends complexity thuscan not be determined beforehand. To achieve high performance, adopt GPU-basedclustering introduced Chiosa Kolb [3] cluster range-intensity images.In context, novel homogeneity criterion handsegmentation system robustly detecting hands, evenunder condition distructingobject, third appears.