Use this URL to cite or link to this record in EThOS:
Title: Advancing radar Doppler techniques for activity monitoring using machine learning
Author: Chen, Qingchao
ISNI:       0000 0004 7964 9653
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This thesis presents research on the design and development of novel machine learning methods for activity monitoring using micro-Doppler signatures (µ-DS) data collected from passive Wi-Fi radar (PWR) and multi-static pulse systems. For PWR, first we propose a phase-sensitive signs-of-life detection method using PWR. Second, we propose a pipeline for µ-DS classification based on the Sparse Representation Classifier (SRC). Third, we adopt and modify the deep transfer network (DTN) to address the limited volume of the dataset. The phase-sensitive method is proved effective in detecting signs-of-life signatures and the modified DTN outperforms shallow methods and the conventional DTN by by 10% and 3% in PWR µ-DS dataset. For active multi-static radar, first we propose Single-Channel (SC-) and MultiChannel (MC-) DopNet for classifying personnel walking with or without a rifle. Based on the Deep Convolution Neural Network, SC-DopNet and MC-DopNet are designed for mono-static and multi-static radar respectively. They have been verified to improve the state-of-the-art results by around 8% in this task. Second, as µ-DS classification depends on two factors of variation: aspect angle and the target personnel, we design two unsupervised deep adaptation networks: i) re-weighted adversarial adaptation network and ii) joint adversarial adaptation network so that they can generalize well to unseen variation factors. The two networks are proved effective to learn useful features invariant to the factors and achieve the state-of-the-art results in both computer vision and µ-DS datasets. Third, we propose cooperative and adversarial relationship between the main activity and the two auxiliary classification tasks, including aspect angle and target personnel classification. To improve the performance of mono-static µ-DS classification, we propose ENet to integrate auxiliary tasks in activity classification network by selecting either the cooperative or adversarial learning strategy. ENet has been verified to outperform the SC-DopNet by 6% in average.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available