Recognition Of Human Action And Identity In Video Sequences

The action recognition is very important topic in computer vision with many of the basic applications such as video surveillance, robots, human interaction with computer, and retrieval of multimedia from Web. Regardless of the input source, the key challenge in human action recognition lies in explicitly or implicitly representing and modeling the temporal evolution of the data, and the other challenge is to know the human identity from low resolution, far distance and no cooperation with the user, like biometric measures such that face recognition, fingerprint and eye iris.Number of published researches in this area continues to increase, but most of these researches are deal only with action recognition without identifying the human who making the action.Most existing papers deal with recognize identity of the pedestrian by measuring step or height measures or additional computational that characterize the human properties, which is based only on walking action.From our best knowledge, very little researches that dealt with recognize different actions and identity at the same time. Due to very difficult to recognize the human face of the remote because far distance of surveillance cameras such as no contacts with the human and takes more computational complexity of time and space.Motivated by these challenges, in this thesis we develop a techniques for human action recognition and identity at the same time, focus on extracting the Spatio-temporal features of the action. We examine and test a novel method for the identification of human identity to binary image and exploit the discriminative power of embedding a watermark techniques as2-D discrete wavelet transform (DWT) that account for the identification of video through the training data, without any complicated calculations and without any additional storage memory, the only post processing is extract a watermark by inverse IDWT. Next, we show that capturing the spatial, temporal and identity of the human that makes action explicitly through dynamical systems leads to an accurate background subtraction of different actions. We also design a novel method than can learn actions from the identity in the web in uncontrolled environment like YouTube and use different methods to extract and learning features and then classified using SVM. Though the coverage of the environment of our methods are still small, such that one person makes one action. The results, including KTH, Weizmann and our lab datasets have been shown that the accuracy of the action recognition and identity is high and better than previous works, it has a great potential as the next generation of learning features in computer vision. In this thesis two parts have been presented.The first part supposed one person makes one action in a controlled environment with homogeneous background and presents the human identity as a watermark embedding as2-D wavelet transform with different methods of action representation and recognition. Each method captures as chapter, which the same identity algorithm and different action modeling and representation.The second part presents an efficient and simple method for learning and recognizes action from identity in the web. The environment is uncontrolled. To identify the human we use Histogram Oriented Gradients (HOG) to detect a human in the video, then training Bayesian classifier for learning features to recognize actions of identity.We review each chapter as follows:The first method uses SIFT to extract local features that deal with video sequences to extract features through a staged filtering that identifies stable points in the scale space.The second method proposes Deep learning, such that feature extraction consists of Deep Belief Network (DBN) on Discrete Fourier Transform (DFT) of tracked features in a video.The third and fourth methods use Chebyshev and Zernike moments respectively, to extract features, because of their properties of orthogonal moments better than geometric moments such is robustness to noise, invariance to translation, rotation, scaling.Finally, the last method proposes learn representations of actions related with identity from the web; use this knowledge to automatically annotate identity in videos.Several experiments have been exploited to explain the effectiveness of our proposed methods. Furthermore content analysis has been presented to prove the efficient-of-our-methods.