Hand Detection And Tracking Based On Depth Maps

Gesture interaction is an important interaction in the human-computer interaction(HCI),by detecting and tracking hands in video images using computers, to understand the intent ofthe controller. Due to its natural, convenient and more suitable for natural human interactionrequirements, and the application scenarios is very extensive, gesture interaction has becomeone of the hotspots of human-computer interaction research.As the traditional RGB camera-based gesture interaction application is rather narrowly,in the paper, we research on hands detection and tracking technology based on3D camera, toimplement a natural gesture interaction. The traditional hands detection and trackingalgorithms are asked the hands are move slowly and micro-deformation, and the trackingalgorithm of tracking and deformation gesture can’t always get rid of the background andlight, with poor robustness. To implement a more natural human-computer interaction, ourresearch objectives is the hands detection and tracking of deformation, rapid movement inmulti-target scenarios.In the part of hand detection, the initial gesture detection can be divided into the methodsbased on static gesture and based on dynamic gesture, in different scenarios. In the paper, weresearch on both the methods. For the detection of static gestures, in order to improve thedetection speed, we present a depth layer segment algorithm based on the depth informationto convert the global scene to local target, and remove the background area, and thencombined with finger detection and DT feature to detection finger and the center of palm. Forthe detection of dynamic gestures, we present a general algorithm based on motionsegmentation, tracking and trajectory identification, to detect specific dynamic gestures in realtime, and then used it in a waving gesture detection.In the part of hand tracking, we present a tracking algorithm based on Camshift,combined with Kalman filter to estimate target location and depth segmentation, to implementtracking of deformable hand, and avoid the interference with other region such as background.For the rapid movement of multi-target tracking, we present a method of combine the motionregion segmentation with the static region segmentation to extract all the movement orstationary target candidate region, and then use our method of optimal matching oftarget-candidate region to locate all the targets, combine with Kalman estimate, at last, useCamshift algorithm to adjust the sizes and positions of all the targets. The method can track multi-objective parallel, improves the real-time and anti-jamming capability of the interactionsystem. Our algorithm optimal matching of target-candidate region also provides the ability torecover tracking when targets overlap and then separated. In addition, to solve the problem ofjudging tracking error and restoration in the traditional tracking algorithm of deformed hand,we improve the Bhattacharyya distance with time weighted to measure the similarity oftracking results and the original target, the method can solve the error detection and rapidrecovery.