An Intellectual Hybrid Learning Approaches for Human Activities Recognition

Abstract

newline ABSTRACT newlineHuman action recognition in videos is one of the most important and active newlinetopics in computer vision, and building more discriminative video newlinerepresentation is crucial for action classification. Vision-based detection newlinebecome ubiquitous in many real-time applications such as surveillance, newlinesecurity, medical healthcare, and so on. The real-time challenge is to detect newlinevideos automatically that contains human movements is known as human newlinebehaviour recognition. newlineTraditional human behaviour recognition algorithms use RGB video as input. newlineThis can be a daunting task due to large fluctuations in behaviour newlinewithin the class, cluttered backgrounds, and possible camera movements newlineand Lighting variations. Recently, the introduction of cost-effective depth newlinecameras offers new possibilities for solving difficult problems. However, it newlinealso introduces new challenges such as noisy depth maps and newlinetime adjustments. newlineIn this thesis, different real-time videos for observing the human behaviour newlinewere effectively analysed during many aspects. Labelled and unsupervised newlinelearning approaches plays a vital role in many sectors due to its high-level newlineflexibility in prediction and classification problems. In similar way, human newlineactivities recognition adopting the learning models to forecast the subject s newlinebehaviour in prior to avoid major health issues. Generally, HAR include many newlineforefront processing such as real-time video observation (i.e., pre-processing), newlinefeature selection, feature optimization in terms of its dimensionality, and newlineclassifying the abnormalities. The proposed learning algorithms followed newlinethese steps in each technique and produced an essential model in forecasting newlinethe human behaviour. newlineAlso, the research proposed a novel feature extractor called Adaptive newlinebackground subtraction Algorithm (ABSA). The outcome of this extractor is newlineskeleton points and silhouettes from the image segments with minimal noise. newlineObserved features are utilized for training and testing purpose of proposed newlineclassifiers. The research also