Depth Motion Map Based Human Action Recognition Using Adaptive Threshold Technique

Nowadays, human action recognition (HAR) has become an emerging research topic for movie understanding, video clip retrieval, human-computer interactions, autonomous driving systems etc. This paper introduces an efficacious feature representation strategy and novel framework for identifying human action by employing depth motion maps (DMMs) based on adaptive thresholding technique. Firstly, each 3D depth frame is projected onto three orthogonal Cartesian planes to form 2D projected maps and concatenating the front view, side view and top view to generate the DMM features.The DMMs capture the motion characteristics from the video sequence. After that, we have adopted adaptive thresholding for separating image objects from video sequence based on pixel intensities so that the radial basis kernel based support vector machine (SVM) are used to classify and recognize the action sequences effitively. We carried out the experiment on MSR-Action3D action dataset and evaluated the classification performance by measuring confusion matrices to provide details and visualization about recognized actions. From the experimental results, we reveal that the proposed framework attains always better performance than other competitive methods on this database.