Sequential learning and shared representation for sensor-based human activity recognition

Hamad, Rebeen Ali (2022) Sequential learning and shared representation for sensor-based human activity recognition. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) hamad.rebeen_phd(19021959).pdf - Submitted Version Download (3MB) | Preview

Abstract

Human activity recognition based on sensor data has rapidly attracted considerable research attention due to its wide range of applications including senior monitoring, rehabilitation, and healthcare. These applications require accurate systems of human activity recognition to track and understand human behaviour. Yet, developing such accurate systems pose critical challenges and struggle to learn from temporal sequential sensor data due to the variations and complexity of human activities. The main challenges of developing human activity recognition are accuracy and robustness due to the diversity and similarity of human activities, skewed distribution of human activities, and also lack of a rich quantity of wellcurated human activity data. This thesis addresses these challenges by developing robust deep sequential learning models to boost the performance of human activity recognition and handle the imbalanced class problems as well as reduce the need for a large amount of annotated data.

This thesis develops a set of new networks specifically designed for the challenges in building better HAR systems compared to the existing methods. First, this thesis proposes robust and sequential deep learning models to accurately recognise human activities and boost the performance of the human activity recognition systems against the current methods from smart home and wearable sensors collected data. The proposed methods integrate convolutional neural networks and different attention mechanisms to efficiently process human activity data and capture significant information for recognising human activities.

Next, the thesis proposes methods to address the imbalanced class problems for human activity recognition systems. Joint learning of sequential deep learning algorithms, i.e., long short-term memory and convolutional neural networks is proposed to boost the performance of human activity recognition, particularly for infrequent human activities. In addition to that, also propose a data-level solution to address imbalanced class problems by extending the synthetic minority over-sampling technique (SMOTE) which we named (iSMOTE) to accurately label the generated synthetic samples. These methods have enhanced the results of the minority human activities and outperformed the current state-of-the-art methods.

In this thesis, sequential deep learning networks are proposed to boost the performance of human activity recognition in addition to reducing the dependency for a rich quantity of well-curated human activity data by transfer learning techniques. A multi-domain learning network is proposed to process data from multi-domains, transfer knowledge across different but related domains of human activities and mitigate isolated learning paradigms using a shared representation. The advantage of the proposed method is firstly to reduce the need and effort for labelled data of the target domain. The proposed network uses the training data of the target domain with restricted size and the full training data of the source domain, yet provided better performance than using the full training data in a single domain setting. Secondly, the proposed method can be used for small datasets. Lastly, the proposed multidomain learning network reduces the training time by rendering a generic model for related domains compared to fitting a model for each domain separately.

In addition, the thesis also proposes a self-supervised model to reduce the need for a considerable amount of annotated human activity data. The self-supervised method is pre-trained on the unlabeled data and fine-tuned on a small amount of labelled data for supervised learning. The proposed self-supervised pre-training network renders human activity representations that are semantically meaningful and provides a good initialization for supervised fine tuning. The developed network enhances the performance of human activity recognition in addition to minimizing the need for a considerable amount of labelled data.

The proposed models are evaluated by multiple public and benchmark datasets of sensorbased human activities and compared with the existing state-of-the-art methods. The experimental results show that the proposed networks boost the performance of human activity recognition systems.

Actions (login required)

View Item

Downloads