Real-Time Multi-Person Action Recognition with a Neural Compute Stick

Deep learning has successfully boosted a performance of action recognition and inspired model developments for it. Specifically, 3D convolutional neural network (CNN) which best fits the purpose of vision-based action recognition has been applied to the task in various forms. In this paper, comprehensive research process for practical multi-person action recognition is presented. We perform various experiments using 3D CNN considering both performance and time efficiency. Distinguished from previous studies, we consider a performance on an embedded platform which consists of an embedded computer, ZED2 camera and a neural compute stick. The neural compute stick has its own memory and can be utilized asynchronously. This is a pioneer work proposing a multi-person action recognition framework using a neural compute stick. Step-by-step experiments verify a validity of the model configuration and the proposed framework.