Based On Joint Points Extraction And Multi-angle Feature Level Fusion Of Human Action Recognition

The human action recognition is a hot issue in the field of machine vision and patternrecognition. It usually refers to identify the action currently being monitored objects byusing its characteristics such as position and posture. In virtual reality, intelligentmonitoring and analysis of sports fields, this method has a wide range of applications.Under the theme of human action recognition in indoor environment, this paper researchesand explores the multi-angle human action recognition, based on in-depth analysis of thehome and abroad research status and recognition algorithm.First of all, this paper researches a new body 2D joint point calibration method.Human silhouette is smooth and connected binary image by image preprocessing. Thenthe skeleton model is extracted by using the method based on the Euclidean distancetransform, establishing body 2D skeleton model whose target area uses pixel-wide unit.Using the key joint 8-neighborhood pixel values to conduct analyze, and combining withthe curvature changes of human lower limb knee point, extracting the actual positioncoordinates corresponding to the human 2D joint points, and then build a human 2D jointpoint model.Secondly, by analyzing the inadequacies of the dynamic time warping algorithm inthe search for a matching path, this paper researches two optimization on DTW algorithmin the case of the recognition rate is not affected, thereby reducing its computationalcomplexity. And then on the basis of the build body joint point model, the optimized DTWalgorithm and the joint point model changes are combined to recognize the single view ofhuman action.Finally, this paper researches and implements a multi-perspective 2D joint pointmodel action recognition method which contains feature level fusion. First, it extracts thejoint points under each feature perspective, normalizes it and constructs initial featurematrix , and then combines feature to form the matrix which has combination feature, thenmakes use of two-dimensional Fisher linear discriminant analysis algorithm to conductfeature compression processing in order to get integrated feature matrix, then combines dynamic time warping algorithm to realize match between pre test integrated set ofcharacteristic matrix set and sample integrated characteristic matrix, thus completes themulti-perspective human action recognition.