Semi-Supervised Learning Approach for Fine Grained Human Hand Action Recognition in Industrial Assembly

Abstract

Until now, it has been impossible to imagine industrial manual assembly without humans due to their flexibility and adaptability. But the assembly process does not always benefit from human intervention. The error-proneness of the assembler due to disturbance, distraction or inattention requires intelligent support of the employee and is ideally suited for deep learning approaches because of the permanently occurring and repetitive data patterns. However, there is the problem that the labels of the data are not always sufficiently available. In this work, a spatio-temporal transformer model approach is used to address the circumstances of few labels in an industrial setting. A pseudo-labeling method from the field of semi-supervised transfer learning is applied for model training, and the entire architecture is adapted to the fine-grained recognition of human hand actions in assembly. This implementation significantly improves the generalization of the model during the training process over different variations of strong and weak classes from the ground truth and proves that it is possible to work with deep learning technologies in an industrial setting, even with few labels. In addition to the main goal of improving the generalization capabilities of the model by using less data during training and exploring different variations of appropriate ground truth and new classes, the recognition capabilities of the model are improved by adding convolution to the temporal embedding layer, which increases the test accuracy by over 5% compared to a similar predecessor model.

Self-Supervised Representation Learning for Robust Fine-Grained Human Hand Action Recognition in Industrial Assembly Lines

“…Subsequently, the suitability for a transfer of the weights to an industrial use case is checked and evaluated on the basis of values such as the amount of labeled data, training time and performance on a test data set. In addition to this extended version of [31], the occurrence of concept drift during fine-tuning is considered and possible responses to prevent performance degradation are outlined. This phenomenon refers to changing circumstances potentially impacting the data and the performance of machine learning models consuming it [9].…”