Constructive, robust and adaptive OS-ELM in human action recognition

We introduce a constructive, robust, and adaptive OS-ELM (Online Sequential Extreme Machine Learning) that combines learning strategies of Constructive Enhancement OSELM and Robust OS-ELM, and add adaptive capability to receive a new target class requirement during sequential learning and applied in human action recognition. The overall strategy is aimed to deal against parameters tuning and new requirements drift problem in the sequential learning process that commonly happened in human action recognition. Our proposed method has an automatic and systematic approach for determining input weight and bias value when the hidden nodes need to be increased to handle a larger training data size and to adjust the output weight when the new target class label is presented during sequential learning. We demonstrated the capability using 2013 Challenge on Multi modal Gesture Recognition open dataset (Chalearn 2013) with uni modal features (skeleton data) only. Our experiments using skeletons of upper body joints features with normalized euclidean distance and projection angle position coordinates to shoulder center and hip center, clustering analysis with k-means for pose based generation and Bag of Pose (BoP) for temporal sequential information. We developed the training sequence scenario to introduce the partial action classes in the initial training and the rest of training data with all classes in the next sequences to simulate the condition when the OS-ELM received new target class requirement drift. Our proposed method gives better accuracy plus adaptive capability compared with SUMO method which is the best uni modal method for chalearn data and still maintain a reasonably small computation time.