A more sophisticated behavioral recognition tool is proposed as I move into a more connected and automated society. The thesis deals with concern that an automatic recognition and actions made by humans in videos and on a real time camera. The automatic understanding of what actions might occur in a human performed videos are known as human action recognition. Because of the numerous problems, like differences in human shape and motion, complex backgrounds, shaking cameras, lighting conditions, and viewpoint variations, these are difficult jobs.
To begin, the most widely used and well-known state-of-the-art approaches are examined, appraised, contrasted, and presented. These strategies are divided into two categories based on the literature review: handcrafted feature-based and deep learningbased approaches. The proposed action recognition framework is then built on top of these deep learning-based methodologies, which are subsequently used throughout the thesis to integrate unique action recognition algorithms in both the handcrafted and deep learning domains.
The knowledge of a deep Convolutional Neural Network model trained on a large-scale annotated dataset to an action detection task with a small training dataset has been experimentally validated. On the same datasets, the comparison study reveals that it outperforms handcrafted feature-based approaches in terms of accuracy. The relative researches also show their higher performance over handcrafted feature-based methods in terms of accuracy on same datasets.
Another ways which are covered on thesis is based on unsupervised deep learning-based approach. This method applies Deep Belief Networks (DBNs) with specific Boltzmann machines for action recognition in unconfined records. The proposed method automatically extracts certain feature representation without any prior knowledge using unsupervised deep learning model. The efficiency of the mentioned method is proved with clear recognition results on an inspiring UCF sports dataset.
Finally, the thesis is made up summarized notes with important experiments and routes in area of human actions by implementing relatively accurate datasets.