Deep learning based classification techniques for human action recognition from multiple foreground video objects

Abstract

Video Foreground Object Segmentation plays a vital role in various newlineapplications such as fraud detection, behaviour analysis. Recognition of a newlinemoving object is a high-level task which can be carried out by extracting the newlineforeground object. This task is difficult due to background characteristics, fast newlinemoving object, ghosting, background modelling and foreground aperture. In newlinethis scenario, object segmentation is achieved by background frame newlinesubtraction, analysing temporally adjacent frames, application of Joint newlineDifference Algorithm (JDA) and a Combination of Different approaches newline(CoD) etc. But, still there exists an inefficiency to segment the dominant newlineobject from a video. Hence, this thesis proposes new techniques for Human newlineActivity Recognition (HAR) by extracting the dominant objects through the newlinegeneration of overlapped regions on the object in the frame using watershed newlineof gradient magnitude. Moreover, the boundary cues are extracted in this newlinework from the Bag of Regions (BoR) for making an object like a region and newlineto restrain the weak boundaries. The proposed object regions are extracted newlinefrom this BoR using Extended Histogram of Oriented Gradients (Ex-HOG). newlineThe Color (L*a*b color space), Texture, Shape (Ex-HOG) features, Color and newlinetexture histogram and their intersection are used in this work for hierarchical newlinesegmentation to segment a dominant object from a video. The multiple cues newlineare used in this work to find the object region. This approach has been tested newlineon a standard benchmark dataset namely Segtrackv2 and the results were newlineanalyzed. newline