Research On Recognition Of Human Action Based On Video Sequence

Recognition of human behavior pattern which is an important bifurcation of Artificial Intelligence has become the research emphasis of Machine Vision area in recent years. Recognition of human behavior pattern contains detection of animated objects from video sequence, the feature extraction from objects and comprehension of these information to recognize human behaviors.Author firstly proposed a novel detection method using vector dependence theory to resolve the disadvantage of images difference and auto-adapt Gauss background modeling respectively. This paper designed a processing procedure to eliminate noises and shape distortion due to shadow effect. After pre-processing, Fourier description from object boundary and skeleton description from object silhouette are extracted. For the Fourier description, this paper eliminates the DC component and normalize the AC components to reduce the effect from excursion, zoom. For the skeleton description, two data structures are designed to store abstract information extracted from the skeleton description, which includes the coordinates of head, arm, chest, crotch, leg and distances between head and chest or arm and chest or leg and crotch and so on. According to these facts, a expert system based on natural language is build to recognize human behavior using predicates deduction. The features extracted from object boundary and silhouette are matched using optimized Euclid distance judgment and expert system based on natural language respectively, that have get primary results.All the algorithms about image recognition in this thesis are programmed with C, and executed in MATLAB. The expert system is constructed with Visual Prolog language. The experiment results show that the database build on Fourier description and skeleton description can work on as fewer as model samples.