Human Gesture Recognition Based On Multi-feature Fusion

The aging in China is becoming a serious problem. Large numbers of old people needsomeone to take care of them. To solve this problem, many countries raise robot for theelderly research to the strategic of view. The key technologies of robot for elderly, such liketarget identification, environment modeling, mission operation and navigation have make bigbreakthrough in recent decades. The human-computer interface problem that the robot facedhas not been solved successfully. In order to find an effectively human-computer interfacemethod that is suitable to robots for the elderly, we proposed a new method ofthree-dimensional human gesture recognition. We use Kinect camera to overcome theproblem that the human-computer interface has low reliability in different lights or withcomplex background to make the robot for the elderly have safe and friendly interactiveawareness.First of all, we analyze the Kinect camera’s features and use it to collect human gesture’sdepth image, color image and skeleton image. And then we build the human gesture databasewith these images. In order to make the following feature extraction not be affected by thebackground, we need to do preprocessing to the images in the database. Before we extract thehuman region from the depth image, we de-noise and smooth the depth image. Then wejoint-calibrated the depth camera and color camera to get the coordinate transformationrelations to map the human region in the depth image to the color image. Now we finished thepreprocessing.Then we use contour characteristic parameters combined with Hu moments, Histogramof Oriented Gradient and skeleton points’ angles to characterize human gesture’s depthinformation, color information and skeleton information. On this basis, we use PrincipalComponentAnalysis to reduce the feature vector’s dimensionality.At last, we combined the Bayesian classifier and boosting algorithm to build ourclassifier. We use Bayesian classifier to classify every kind of human gesture information andthen use boosting algorithm to combine these classification result. And now we finished thehuman gesture identification. On this basis, we use our classifier to test the samples in thedatabase, non-specific samples and samples in different light conditions, verifying theeffectiveness of our article’s recognition algorithm. At the same time, it also proves that ourrecognition algorithm solved the technical difficulties of user independence and perspectiveindependence, and can identify the human gesture in general environment.