HMM Based Dynamic Hand Gesture Recognition

Vision-based hand gesture recognition is an important research topic in the field ofhuman-computer interaction. If people use natural gestures to replace the intermediary whenthey interact with computers, it will be more natural and intuitive, and closer to humancommunication habits. In addition, the research of gesture recognition has many practicalapplications.How to segment pre-defined operating gestures from the continuous sequence of handgestures is considered as a highly difficult problem and research priority in the research ofhand gesture recognition. At present, there are two major difficulties: one is the variation ofgesture in spatio-temporal, when a hand gesture is repeated by a user or did by different users,the hand shape, velocity and margin of each hand motion often varies, which cause the samegesture differs in trajectory and duration; the other difficulty is how to segment and extractisolated gestures from continuous gestures, a continuous gesture sequence containsmeaningful gestures and atypical gestures, we need to determine the start and the end pointsof meaningful gestures that are embedded in the continuous gesture sequence, automaticallydivide the sequence into isolated gestures, and resolve the issue of return stroke andtransitions etc. To overcome these problems, HMM is used in this paper, which can analysistime series under the spatio-temporal variation，is capable of modeling complex gestureseffectively and can rejection atypical gestures.In the paper, we mainly did the work of the following three parts based on the existingdynamic gesture recognition algorithms:Firstly, we researched the isolated gesture recognition based on Hidden Markov Model,and improved the training algorithm, to improve the recognition rate, our method trained thefalse recognized gesture samples again; To solve the problem that the typical model can notrefuse atypical gestures, we proposed a method to extract threshold using the training samples.With our algorithm in this paper, the recognition rate of pre-defined gestures achieved98.06%,and the reject rate of undefined gestures achieved91.25%.Secondly, to get an adaptive threshold for the online recognition of dynamic gestures, weresearched the threshold model-based gesture recognition algorithm. With the proposedalgorithm, we got a recognition rate of97.87%for pre-defined gestures and reject rate of91.25%for undefined gestures. In addition, we analyzed the advantages and disadvantages of threshold model in continuous gesture sequence recognition.Finally, we proposed an inflection points detection algorithm, to detect the candidatestarting points and ending points in the gesture sequence, and then segmented candidategesture sequences by the inflection points. By combining with the isolated gesture recognition,we implemented a dynamic gesture online segmentation and recognition framework. To solvethe problem of segmenting complete sequence and recognizing the embedded gesture, weimproved the above framework: when a pre-defined gesture is detected, the algorithm doesnot output it immediately, but detects the end point of the detected gesture, and if it is anembedded gesture, a recursive algorithm will be started to detect the longer gestures, to makesure the output is a complete sequence.