Font Size: a A A

Human Action Recognition Using Support Vector Machine

Posted on:2013-06-04Degree:MasterType:Thesis
Country:ChinaCandidate:HEBABLIA NacirFull Text:PDF
GTID:2248330377959330Subject:Control theory and control engineering
Abstract/Summary:PDF Full Text Request
Action recognition has become a very important topic in computer vision, with manyfundamental applications, in robotics; video surveillance; human computer interaction; andmultimedia retrieval among others and a large variety of approaches have been described.The purpose of this thesis is to give a description of the action that was beingperformed by a human in video. A Weizmann database was adopted here to achieve this workwhich comprises videos with a homogeneous and static background. It consists in ten actionclasses: bending, running, walking, skipping, jacking, jumping forward, jumping in place,galloping sideways, waving with two hands, and waving with one hand. Each video containsa single periodic action, performed by a single person.First, we will do some pre-processing methods in order to make the main task of thethesis easier after. Those pre-processing methods are mainly shown up in three small tasks:moving target detection; morphological processing and feature extraction. The moving targetin the scene is detected by subtracting the estimated background image from the imagesequences.The estimated background image is gotten by one of many powerful methods used inthis domain which is the background difference method. This latter in order to be achievedneeds two other small steps: background modeling and background subtraction.There are so many methods used to give an estimation or a model for the static objectsthat make the background image; such as: Running Gaussian Average; Temporal MedianFilter; Mixture of Gaussians; Kernel Density Estimation; Sequential Kernel DensityApproximation; Cooccurrence of Image Variations; Eigenbackgrounds and so forth…In this thesis, the Running Gaussian Average method was adopted because of itssimplicity, better results; not time-consuming and good for implementation.Once the estimated background step is done, the foreground images will be easy to getand they are done just by subtracting off the background image from the current frames withinthe video.Then; a step named noise reduction and morphological processing will have birth inorder to enhance the shape of the detected human silhouette by filling holes and some partsmissing within it. This step is based on dilation, erosion, opening and closing operations thatare the four principals. After that; in the step feature extraction; local (corner detection using Harris cornerdetector) and global (contour detection) features extraction are studied respectively. Finally,we adopt the Support Vector Machine (SVM) as the classifier algorithm which is a concept instatistics and computer science for a set of related supervised learning methods that analyzedata and recognize patterns, used for classification and regression analysis.
Keywords/Search Tags:Human action/activity recognition, background subtraction, corner detection, (2D)2PCA, SVM, Feature extraction, computer vision, Survey
PDF Full Text Request
Related items