Human Assembly Task Recognition in Human-Robot Collaboration based on 3D CNN

We focus on the problem of high-precision and long-timespan human assembly task recognition from videos in human-robot collaboration area. Toward this 3D Convolutional Neural Networks (3D CNN) model is used in this paper. 3D CNN has been proved behaving well in human action recognition. But, when it comes to human assembly task recognition, there are two new problems. Firstly, human assembly task recognition needs higher recognition precision as the background of the assembly station never changes and the assembled products is always the same one. Secondly, recognizing an assembly task needs much longer timespans than just recognizing an action as a task always consists of many actions. Aiming at the first problem, we augment the diversity of the data by rotation and mirroring process. To solve the second problem, we increase the number of input frames from 10 to 100. Experiment result shows that the recognition accuracy finally reaches to 0.82.