An Accurate Random Forest-Based Action Recognition Technique Using only Velocity and Landmarks’ Distances

Human action recognition has gained significant attention recently as it can be adopted for various potential applications, notably for smart activity monitoring for surveillance and assisted living purposes. However, recognizing human activities is a challenging task because of the variety of human actions in daily life. In this work, we propose to extract skeletons from RGB images using deep learning architectures while adopting hand-crafted features for action recognition. The main idea is to extract few features from the skeleton data in order to ensure low feature dimensionality while leading to fast and accurate action recognition. In fact, our objective is to define hand-crafted features that are clearly explainable and related to the kinematics. Each joint is modeled by a single feature vector that encodes only the most essential kinematic information to characterize an action: the relative locations of the joints as well as the velocity of these joints. We have validated the proposed technique on the challenging UTD-MHAD multimodal action dataset and the preliminary obtained results are very encouraging.