Research On Gesture And Posture Recognition In Human Computer Interaction System

With the development of science and technology, the third generationhuman-computer interaction systems are booming. The interaction systems based ongestures and postures are more friendly and convenient, so algorithms based on gesturesand postures recognition are becoming more important. But the image recognitionalgorithms still have some problems, such as calculation is complex, the recognitionrate is not ideal, etc. To solve these problems, this dissertation researches gestures andpostures recognition by the usage of visible RGB image and the depth image. The maincontributions are as follows:1. In the visible gesture recognition system, the existing algorithms’ performanceor speed is not very ideal. The author first collects a number of positive samples,calibrates the samples, adds negative samples, extracts HOG feature, and finally trainsthe weak classifiers with AdaBoost algorithm to a strong cascade classifier to achievethe hand gesture detection. Then the author uses several common gestures with HOGfeatures plus tree classification to recognize common gestures. In order to run faster, thedissertation applies the method of accelerated processing and the correspondingoptimization calculation..2. For complex background of visible RGB images, the recognition performancedecreased. But it will be more stable if using depth images as input. The authorproposes a novel algorithm which recognizes static hand gestures on depth images using3D shape context feature, and improves the algorithm by arm major axis correction andcontour-center sampling. Arm major axis correction can solve the rotation and thesampling method increases the recognition rate. Besides, the directly usage of the3Dspace information makes the algorithm be scale and distance invariance. Theexperiment result shows that the average recognition rate is better than existingalgorithms. The recognition result can be used for follow-on human-computerinteraction operations.3. The randomized forests training algorithm used in depth image human poseestimation cause problems of large resources and training time cost. The dissertationproposes the parallelization design using Message Passing Interface on small cluster server, then optimizes to decrease the storage and bandwidth cost. Experiments showthat the processing speed is enhanced by about30times, and every training cost is lessthan one day on the cluster. The proposed parallelization method can finish the trainingin time to adjust and test the parameters.4. In predicting3D positions of human skeleton joints, the errors of some jointsmay be large. The dissertation proposes a scheme to decrease the cost of training andadd post-processing using multi-view depth images to improve the performance. Theauthor first generates small image sets which contain important and typical humanposes, then trains the classifier using a small cluster in the laboratory and estimates thehuman skeleton joints. For the joints with low confidence, the post-processing algorithmis added using the information of side view and top view depth images which couldimprove the precision of skeleton joints obviously. The experiments using relativelysmall training sets show that average error of skeleton joints is less than the originalalgorithm.