Human activity recognition using temporal convolutional neural network architecture

In health care and other fields, the detection and recognition of human actions or activities are essential in the context of human–robot interaction. During the last decade, many approaches for human activity recognition have taken advantage of high-performance computing devices. These devices make use of various sensors and improve the quality and efficiency of the results. With the aim of using a non-invasive method, we propose the design of a temporal convolutional neural network that uses spatio-temporal features to analyze and recognize human activities using only a short video as input. The proposed architecture is based on a 3D convolutional layer and a convolutional long short-term memory layer. Our methodology leverages the time-motion features with the spatial location of the activities performed by people to improve the accuracy of the classification results. This design makes optimal use of computational resources to achieve training/classification in a short period of time, and consequently, obtain real-time classification results. The computer simulations showed that our method provided superior state-of-the-art classification results for human activities even for those methods that require information from more sensors.