Pose-based approach for human action recognition

In the recent times, human action recognition has been active research area in computer vision research due to the increase of security threats like terrorism and proliferation of advanced technologies used between the interaction of human and electronic devices. The objective of this project is to look at the possibility of using pose-based approach for human action recognition, and examine its usability and performance using various methodologies. Multi-view of human action is designed with two main methods. Firstly, in the action representation method, the human body is used as a posture vectors to train the Self-Organizing Map (SOM), which in turn, produces a topographic map. This map shows the key poses that represent the entire action videos. Next, the posture vectors are mapped into the Fuzzy Membership Vectors to produce the action videos. Each action video is a feature of an action. In the second method, action classification, different classifiers like Backpropagation Training Algorithm, Extreme Learning Machine (ELM) and Naïve Bayes Classifier were chosen to classify the action classes. Backpropagation Training Algorithm was chosen because of its ability to solve non-linearly separable problems. To achieve the concurrent objectives of short computing time and high accuracy in multi-class classification, ELM was chosen. Lastly, in the Naïve Bayes classifier, a tiny portion of data is required to approximate the mean and variance of the variable needed for classification. The robustness of this aforementioned system was tested on 4 events from WVU multi-view Action Recognition datasets and 10 actions from single view Weizmann datasets. This system was subsequently compared with the power spectrum feature technique. It was shown that human action recognition using posed-based approach proved superior to power spectrum feature technique.Bachelor of Engineerin