A field research on human action recognition approaches with video data sets

Human action recognition is the prediction of people's movements using machine learning and deep learning methods. With the increasing interest in machine learning and deep learning methods in recent years, the issue of detection of human movements has also been developing. With the development of technology, action detections, which were first extracted from static images, started to be performed on videos and even in live streams. Human action recognition methods are one of the learning-based methods we have started to see in daily life. Prediction of thieves, criminals or individuals who will carry out dangerous activities has started to be actively used in traffic predictions of pedestrian activities and in many other areas. Furthermore, with the development of learning methods on the subject of human action recognition, the areas of use are expanding with the increase of speed and correct recognition performances and the development of practical methods. In this study, fifteen different articles drawing attention with their different methods on Human action recognition are discussed and a research is prepared by examining each one in detail. In this research, approaches that give a new perspective to human action recognition in terms of methods will be examined. All the articles discussed throughout the research carry out recognition on the videos. All these recognition 9 processes are categorized by a taxonomy and five main categories are created. These categories are; network-based approaches, motion-based approaches, multiple instance learning based approaches, dictionary-based approaches and histogram- based approaches. All articles were examined according to this categorization. Explanation of the methods for each of the articles examined, the stages of development, the need arising from it, the studies on the data sets and the accuracy results obtained were detailed. In addition, datasets containing videos from daily life, vehicles in traffic, sports and drone videos were also considered and a comparison table was created by examining each dataset in order to gain an idea for new researchers. Tables containing training and test separations of articles using data sets are also included. This research will be a reference to future studies of all articles and data sets detailed.