Leveraging key-points Motion History Maps for Human Activity Recognition

There has been a great interest in investigating machine learning techniques for rec-ognizing human activities from video sequences or images. The task is particularlychallenging due to potential problems that could appear related to clutter, partial oc-clusion, changes in scale, viewpoint, lighting, and appearance. Additional, many fieldsof applications could require an activity recognition tool including video surveillancesystems, human-computer interaction, e-health applications, and robotics. In this pa-per, we present a novel two-phase approach for extracting and learning knowledgefrom a subset of possible human actions, working in a multi person scenario. Ourapproach first employs a deep neural network to perform the pose estimation and theskeleton key-points extraction of a subject. Successively, the key points are used tofirstly track each subject in the scene and secondly build the motion history images.These ones represent the inputs of a convolutional neural network able to recognizeand correctly classify seven different possible human actions. Despite every sub-blockof our machine learning pipeline being an off-the-shelf solution by itself, the wholesystem represents a pretty impactful innovation. In fact, it enables fast and efficientreal-time multi-target activity recognition, which is something seldomly investigatedin literature, but of great importance for real world applications. Experimental evalu-ations demonstrate that the solution is competitive with existing approaches, reachingan accuracy near to 90% on a real world-dataset.