Mhad: Multi-human action dataset

Abstract

This paper presents a framework for a multi-action recognition method. In this framework, we introduce a new approach to detect and recognize the action of several persons within one scene. Also, considering the scarcity of related data, we provide a new data set involving many persons performing different actions in the same video. Our multi-action recognition method is based on a three-dimensional convolution neural network, and it involves a preprocessing phase to prepare the data to be recognized using the 3DCNN model. The new representation of data consists in extracting each person’s sequence during its presence in the scene. Then, we analyze each sequence to detect the actions in it. The experimental results proved to be accurate, efficient, and robust in real-time multi-human action recognition.