Machine Learning for Video Action Recognition: a Computer Vision Approach

The automatic detection of video action is still a challenging research task. In this paper, we consider a first atomic approach and its empirical evaluation to classify a single action in a short video sequence based on DITEC image characterization method. The presented method combines four different concepts: global image descriptors, image transformation algorithms, Machine Learning paradigms for supervised classification and Feature Subset Selection (FSS) techniques. Using DITEC descriptors, which are based on the Trace Transform, the information contained in a video is handled as an image. This allows us to apply Image Processing solutions for the analysis of the video, more concretely, of the occurring action. Key features are extracted to nourish Machine Learning classifiers in order to predict the action. The final step is to use a Feature Subset Selection (FSS) standard method to select the most accurate attributes for the identification of the action. The idea of understanding videos as images widens the possibilities for the analysis of temporal behaviour of actions within a video.