Study On Hand Gesture Recognition And Interaction Parameters For Freehand Human-Computer Interaction

In recent years,Human-Computer Interaction(HCI)based on image processing,which does not require users to touch any input device,has become a novel and promising way to interact with computer applications.Compared to traditional interaction methods based on mouse and keyboard,HCI based on image processing possesses such advantages as being natural and immersive,and could be better suited for virtual reality systems or public exhibi-tion systems.There are two key techniques with the realization of HCI based on image pro-cessing,one being hand gesture recognition which make computers able to keep track of and recognize the actions of user's hand,and another being interaction mapping which translates the hand actions to commands for computer.Both techniques are important for user experi-ence in HCI based on image processing.Currently the state-of-the-art hand gesture recogni-tion algorithms suffer from such problems as low recognition accuracy,slow recognition speed,etc.,and interaction mapping is defined differently by each HCI developer.In order to solve such problems,this paper studies the two key techniques,and the achievement and innovations are:1)A new method to improve the performance of feature-parameterized random forest algorithm as well as ways to implement random forest algorithm to make it extraordinarily fast are proposed.A particle swarm optimization process is introduced into the training pro-cedure of each node of random decision trees in the random forest,which takes the place of the traditional process of randomly selecting feature parameters,and thus makes the splitting function on each node locally optimized.The result of experiments shows that the improved algorithm is more efficient in classifying compared to the traditional algorithm,i.e.lower error rate on the same level of tree using either ideal or unideal datasets.Designs on data structures,way to index samples,way to parallelize the algorithm and many other aspects of the implementation of random forest algorithm are also proposed to make the algorithm run extraordinarily fast.2)A novel hand posture recognition algorithm based on feature-parameterized random forest is proposed.Scale and in-plane rotation invariant features are designed,which are ca-pable of embracing both hand contour information and depth texture information.A hand posture recognition algorithm is proposed,which adopts feature-parameterized random forest to find combinations of the designed features to effectively classify hand postures from dif-ferent view angles.The result of experiments shows that the recognition rate of the proposed algorithm is very close to that of the state-of-the-art algorithms,while being superior on pos-tures that are highly similar to each other.The speed of the proposed algorithm is about 600 frames per second,making potential real-time applications running on low-end processing units possible.A real-time demo of the algorithm is also developed to demonstrate its practi-cal effectiveness.3)A 3D virtual environment freehand navigation application for a remote meeting sys-tem is designed,and a method to study the effect of different navigation parameters on user experience is proposed.A freehand navigation application based on the proposed hand pos-ture recognition algorithm and a new hand detection algorithm is designed for a remote meet-ing system.Three parameters of the speed mapping function for freehand navigation are stud-ied,and experiments on 16 subjects conclude the effect of different parameter values to effi-ciency of navigation,user preferences and user fatigue.The conclusions could provide rea-sons for the design of navigation mapping functions in future 3D virtual environment free-hand navigation applications.