Study On Real-time Unmarked Hand Tracking

With the rapid development of virtual reality, robotics and artificial intelligence, the diversified interaction based on human body is becoming the new direction of the development of human-computer interaction.Human hand can deliver rich interactive information, therefore, the hand-based interactive technology has gradually become a hot spot and the demand for hand tracking research is increasingly urgent. The high degree of hand structure freedom and the flexibility of movement make the technology vulnerable to self-occlusion, self-intersection and other issues. Referring to the calculation on hand motion parameters of the monocular depth image, research on unmarked real-time motion tracking is carried on. In this dissertation, hand tracking algorithm based on hybrid method has been designed, solves the problem of the extraction of the region of interest in complicated background, and improves the premature convergence and low precision of model fitting. This method achieves high image segmentation accuracy, and the hand tracking based on hybrid method can also achieve certain effects. The main contents of this dissertation include the following aspects:Aiming at the extraction of the monocular depth image in the complex background, a hand image segmentation algorithm based on randomized forest classifier has been studied. By combining the characteristics of the threshold segmentation method, a difference feature based on the depth of the hand centroid is introduced. The image feature extraction algorithm is improved to minimize the depth of difference between the applications of different scenes. Finally, median filter and morphological open algorithm are used to remove the camera noise and fill the holes. The results of experiments show that the method proposed in this dissertation is easier to implement than the existing methods and has strong generalization ability for unknown samples.The classification of hand gestures based on randomized forest classifier has been studied on the basis of discriminative tracking method.Firstly, the forward driving and parameter reverse of the model are studied by Ogre on three-dimensional model. Then, a synthetic depth map generation method with precision approximation is designed. Next, based on the purpose of hand gesture classification, the feature extraction algorithm is improved on the basis of depth difference feature. Finally, a randomize forest algorithm based on information gain and distribution histogram is implemented. The experiments show that the algorithm has high accuracy on the test data set, and can achieve high accuracy for the real set which are generated from a certain range.The model fitting technique of particle swarm optimization algorithm has been studied in this dissertation. To begin with, the particle update and initialization criteria are improved by the idea of hybridization and mutation in genetic algorithm. Then, in order to evaluate the prediction parameters accurately, a normalized method of rendering image based on camera focus is designed considering the perspective relationship between model depth and image resolution. Aiming at the problem of computational efficiency, this dissertation designs a PSO parallel computing method based on CUDA and realizes the mixed programming technology of Ogre and CUDA.In order to verify the effectiveness and feasibility of the tracking algorithm proposed in this dissertation, an experiment on unmarked real-time hand tracking is carried out for one-hand motion tracking as example. Experimental results are obtained to meet the expected target and the effectiveness of the tracking algorithm has been verified.