Dynamic Depth Data Matching And Its Application

The depth information means the distance from the camera lens to the target object.Dynamic depth data, namely several depth frames associated with chronological, offersseveral advantages over traditional intensity video for the expression of the object’saction, such as working in low light levels, giving a calibrated scale estimate, beingcolor and texture invariant, and resolving silhouette ambiguities in pose. With thepopularity of Kinect somatosensory equipment, more and more dynamic depth data willappear on the Internet combined with optical video and becoming one of the mainmedia in the future. In the past, content based video matching and retrieval often In thispaper, we pay attention to the study of dynamic depth data matching, and specificstudies are as follows:First, we capture the dynamic depth dataset. For the experiment in this paper, wecapture a database that consists of17kinds of human actions from10persons ofdifferent gender who have different height ranged from158cm to182cm. These actionsreflect the superiority of the dynamic depth data over the optical video for theexpression of the object’s action and also reflect the superiority of our matching methodover the skeleton information.Second, we study the dynamic depth data matching method. A new3D shapecontext descriptor is proposed for the matching of static depth image, which is proved tobe very effective and robust in the expression of human pose, and then we apply theDynamic Time Warping (DTW) to the dynamic depth data matching. Theexperiment results show that using3D shape context&&DTW as the matching methodof the dynamic depth data obtains remarkable matching accuracy. Finally, for thematching efficiency problem, we use Bag of Word model to accelerate the matchingprocess. Specifically, each piece of static depth image is quantized into visual wordspackage, so that the original multi-feature matching problem are simplified intotwo histogram matching problem. From large number of experiments, we verified theeffectiveness of our matching algorithm, and we can also see our accelerating strategyobtain a high matching efficiency while also maintaining the matching accuracy.Finally, we apply our dynamic depth data matching strategy to develop a dynamicdepth data retrieval system. We try to retrieve the same action in our dynamic depthdataset when users completed one of our predefined17actions in front of MicrosoftKinect. User might get the retrieval result within a very short time (within6seconds).Practical use shows that our system has high retrieval efficiency and accuracy.