Transfer Learning for Human Activity Recognition in Logistics

在工业4.0时代，人工活动在物流领域仍然占主导地位。关于相关人类活动的发生和持续时间的详细信息对于仓储效率以及整个供应链至关重要。由于人工评估在经济上是不合适的，人类活动识别（HAR）的方法获得相关性。HAR是一种分类任务，用于从时间序列中识别人体运动，已用于智能家居、康复和健康支持等应用。基于非侵入性和高度可靠的体上设备的HAR具有特殊意义，因为这些设备在具有挑战性的场景中扩展了其潜力。训练一个分类器需要大量的数据，因为人类的运动是高度可变和多样化的，特别是在物流部门的不同环境中。本项目的目标是开发一种方法，以避免为物流中的HAR创建和注释高质量的身体设备数据的巨大努力。不同的物流场景将在配备高精度光学运动捕获（oMoCap）的参考场中复制。在这种约束环境中，oMoCap和身体上的设备数据将被同步捕获。所有场景的组合oMocap记录构成参考数据集。转移学习和零触发学习的方法将能够在各种场景中构建这样的参考数据集。机器学习方法，特别是深度学习将被考虑用于处理时间序列和基于参考oMoCap-dataset训练分类器。分类器将允许对同步的体上设备数据进行自动注释。此外，将考虑从原始oMoCap数据创建额外合成数据的方法。将通过将其与来自真实的仓库的手动注释数据进行比较来检查在自动注释和合成的身体上设备数据上训练的分类器的性能。

In the age of the Industry 4.0, manual activities remain dominant in the logistics sector. Detailed information on the occurrence and duration of relevant human activities is crucial for warehousing efficiency and thus the entire supply chain. As manual assessment is economically inexpedient, methods of human activity recognition (HAR) gain relevance. HAR is a classification task for recognizing human movements from time-series that is already used in applications such as smart-homes, rehabilitation and health support. HAR based on non-invasive and highly reliable on-body devices is of special relevance, as these devices extends its potential in challenging scenarios. Training a classifier demands a large amount of data, as human movements are highly variable and diverse, in particular in the diverse environments of the logistics sector.The objective of this project is to develop a method for avoiding the tremendous effort for creating and annotating high quality on-body-devices data for HAR in logistics. Different logistics scenarios will be replicated in a reference field that is equipped with a highly accurate, optical-motion capturing (oMoCap). In this constraint environment, oMoCap and on-body device data will be captured synchronously. The combined oMocap recordings of all scenarios constitute a reference dataset. Methods of transfer- and zero-shot learning will enable to constitute such reference dataset across the scenarios. Methods of machine learning, especially, deep learning will be considered for processing time-series and for training a classifier based on the reference oMoCap-dataset. The classifier will allow for an automated annotation of the synchronized on-body devices data. Furthermore, methods for creating additional synthetic data from raw-oMoCap data will be considered. The performance of the classifier that is trained on the automatically annotated and synthetic on-body device data will be examined by comparing it to manually annotated data from a real warehouse.