Human Action Recognition Using 3D-Convolution Neural Networks

A keen analysis of human activities in the realistic environment provides vast areas like surveillance systems,customers understanding,shopping attitude,normal and abnormal actions to researchers with large amount of applications.However,finding the accurate recog?nition of actions is a challenging task due to various limitations such as cluttered backgrounds,occlusions,viewpoint variations,etc.These limitations about the automatic recognizing of human actions in videos must be kept in mind.Real-time automatic recognition for human action recognition(HAR)and uncontrolled video information,like"surveillance videos" is the main focus of this research.Recently,the researchers have tried to improve the accuracy and precision of video-based recognition systems,but have not really considered the efficiency of the recognition systems.An efficient system with highly precise values is the main consideration of this study.This thesis also focuses on the instrumentation of recognition in real-time environment.Moreover,it is more crucial to recognize and analyze human actions in a complex environment.The aim of this research is also to differentiate between normal and abnormal actions,which followed through a systematic way for classification.Comprehensive studies show that recently implemented classifications are based on complexity,handcrafted features extracted by raw inputs.The Convolution neural networks have the ability of performing direct action on the raw inputs but also have limitations of handling 2D inputs.The research introduces a novel 3D-Convolution Neural network for human action recognition.Moreover,the proposed method is a fully automated deep model for human action recognition.The learning process is without any prior knowledge to classify human actions.The proposed method consists of two steps:First,applying of a 3D-Convolution neural network for extraction of spatio-temporal features for further process.Second,training of a Recurrent Neural Network(RNN)with one layer for classification in a sequence of the temporal evolution of the learned features for each time step.The proposed method shows the experimental results on the KTH and UCF11 datasets,which provides us with tremendous results showing its high efficiency.