New Summit website coming in May 2021!

                   Check the SFU library website for updates.

Engineering Science - Theses, Dissertations, and other Required Graduate Degree Essays

Receive updates for this collection

Autonomous task-based grasping for mobile manipulators

Author: 
Date created: 
2021-02-12
Abstract: 

A fully integrated grasping system for a mobile manipulator to grasp an unknown object of interest (OI) in an unknown environment is presented. The system autonomously scans its environment, models the OI, plans and executes a grasp, while taking into account base pose uncertainty and obstacles in its way to reach the object. Due to inherent line of sight limitations in sensing, a single scan of the OI often does not reveal enough information to complete grasp analysis; as a result, our system autonomously builds a model of an object via multiple scans from different locations until a grasp can be performed. A volumetric next-best-view (NBV) algorithm is used to model an arbitrary object and terminates modelling when grasp poses are discovered on a partially observed object. Two key sets of experiments are presented: i) modelling and registration error in the OI point cloud model is reduced by selecting viewpoints with more scan overlap, and ii) model construction and grasps are successfully achieved while experiencing base pose uncertainty. A generalized algorithm is presented to discover grasp pose solutions for multiple grasp types for a multi-fingered mechanical gripper using sensed point clouds. The algorithm introduces two key ideas: 1) a histogram of finger contact normals is used to represent a grasp “shape” to guide a gripper orientation search in a histogram of object(s) surface normals, and 2) voxel grid representations of gripper and object(s) are cross-correlated to match finger contact points, i.e. grasp “size”, to discover a grasp pose. Constraints, such as collisions with neighbouring objects, are incorporated in the cross-correlation computation. Simulations and preliminary experiments show that 1) grasp poses for three grasp types are found in near real-time, 2) grasp pose solutions are consistent with respect to voxel resolution changes for both partial and complete point cloud scans, 3) a planned grasp pose is executed with a mechanical gripper, and 4) grasp overlap is presented as a feature to identify regions on a partial object model ideal for object transfer or securing an object.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Kamal Gupta
Mehran Mehrandezh
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) Ph.D.

Deep learning for optical coherence tomography angiography: Quantifying microvascular changes in diabetic retinopathy

Author: 
Date created: 
2020-12-18
Abstract: 

Optical Coherence Tomography Angiography (OCT-A) permits visualization of the changes to the retinal circulation due to diabetic retinopathy (DR), a microvascular complication of diabetes. Machine learning applications have directly benefited ophthalmology, leveraging large amounts of data to create frameworks to aid clinical decision-making. In this thesis, several techniques to quantify the retinal microvasculature are explored. First, high-quality, averaged, 6x6mm OCT-A enface images are used to produce manual segmentations for the corresponding lower-quality, single-frame images to produce more training data. Using transfer learning, the resulting convolutional neural network (CNN) segmented the superficial capillary plexus and deep vascular complex with performance exceeding inter-rater comparisons. Next, a federated learning framework was designed to allow for collaborative training by multiple participants on a de-centralized data corpus. When trained for microvasculature segmentation, the framework achieved comparable performance to a CNN trained on a fully-centralized dataset.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Marinko Sarunic
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

Downsized flexible capacitive strain sensing filament and its application in knee angle tracking during walking

Author: 
Date created: 
2019-12-18
Abstract: 

Assistive Rehabilitation technologies have been a prominent research topic in recent years for patients who suffer from impaired lower extremity motor abilities. Continuous tracking of knee movement can provide valuable insight into the effectiveness of therapeutic or surgical interventions such as improvements in range of motion or stability over the course of the treatments. Video-based motion tracking is an industry standard for motion tracking, but its usage is limited in a clinical setting due to its prohibitive cost and space requirement. This study proposes a downsized flexible capacitive strain sensing filament that can be weaved into textiles to achieve in situ motion tracking. Its effectiveness is shown in a knee joint angle tracking with video-based motion capture as reference. Sensor-predicted knee angle is 99% accurate when compared to the reference with root mean square error of 1.79 degrees. An improved sensor is also fabricated and characterized to show enhanced performance.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Carlo Menon
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

Super learner implementation in corrosion rate prediction

Author: 
Date created: 
2021-02-22
Abstract: 

This thesis proposes a new machine learning model for predicting the corrosion rate of 3C steel in seawater. The corrosion rate of a material depends not just on the nature of the material but also on the material's environmental conditions. The proposed machine learning model comes with a selection framework based on the hyperparameter optimization method and a performance evaluation metric to determine the models that qualify for further implementation in the proposed models’ ensembles architecture. The major aim of the selection framework is to select the least number of models that will fit efficiently (while already hyperparameter-optimized) into the architecture of the proposed model. Subsequently, the proposed predictive model is fitted on some portion of a dataset generated from an experiment on corrosion rate in five different seawater conditions. The remaining portion of this dataset is implemented in estimating the corrosion rate. Furthermore, the performance of the proposed models’ predictions was evaluated using three major performance evaluation metrics. These metrics were also used to evaluate the performance of two hyperparameter-optimized models (Smart Firefly Algorithm and Least Squares Support Vector Regression (SFA-LSSVR) and Support Vector Regression integrating Leave Out One Cross-Validation (SVR-LOOCV)) to facilitate their comparison with the proposed predictive model and its constituent models. The test results show that the proposed model performs slightly below the SFA-LSSVR model and above the SVR-LOOCV model by an RMSE score difference of 0.305 and RMSE score of 0.792. Despite its poor performance against the SFA-LSSVR model, the super learner model outperforms both hyperparameter-optimized models in the utilization of memory and computation time (graphically presented in this thesis).

Document type: 
Thesis
File(s): 
Supervisor(s): 
Bonnie Gray
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

PPM level gaseous ammonia detection using laser Induced fluorescence on vapochromic coordination polymers

Author: 
Date created: 
2020-09-04
Abstract: 

The detection of ammonia in parts per millions range has been challenging in sensors research and is of great importance for industrial applications. This thesis document efforts to develop and test a low-cost optical detection system for ppm level ammonia measurements utilizing a Vapochromic Coordination Polymer (VCP) Zn[Au(CN)2]2 as the sensing material. Upon high concentration ammonia exposure, the polymer’s fluorescent peak under near-UV stimulation undergoes a spectral shift from 470nm to 530nm, while the intensity increases by 3~4X. At ammonia concentrations < 1000ppm, the spectral shift becomes hidden within the overall changing fluorescent spectrum shape so simple detection methods do not work. The key point in this analysis is to note the way the spectrum changes in each wavelength bins varies in different ammonia concentration exposures. We then developed two customized spectral processing techniques named Spectral Region Subtraction (SRS) method and Sum of Integrated Emissions (SIE) method to characterize hidden changes in spectral shape for concentrations < 1000ppm. Both methods give excellent sensitivity between 0 – 50 ppm and > 300 ppm. For wide-range concentration detection, a combination of two metrics have to be used together.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Glenn Chapman
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

A deep learning approach for detecting epileptic spike in magnetoencephalography signals

Author: 
Date created: 
2020-07-16
Abstract: 

Epilepsy is one of the most serious neurological disorders that affects people of all ages. In Canada, an average of 15,500 people discover epilepsy symptoms each year [1]. Numerous scholars have conducted extensive research in automated detection of epilepsy spike for presurgical assessment. However, the study of Magnetoencephalography (MEG) spike detection is limited to under 30 patients’ data. In this thesis, we explore a deep learning approach for detecting spike in interictal MEG recordings of up to 300 epileptic patients in an automated fashion. We evaluate the convolutional neural network architecture and long short-term memory method on both 2D images and 3D spatiotemporal MEG recordings. For 2D images, we tested a simple 3 layer Convolutional Neural Network (CNN/ConvNet) model and a transfer learning model, and achieved an accuracy of 83.12% and 82.73%, sensitivity of 91.66% and 78.52%, and specificity of 74.58% and 86.94%. For 3D spatiotemporal data, we tested the 3 dimensional CNN model and Long short-term memory (LSTM) model to get 86.04% and 83.09% in accuracy, 92.37% and 87.18% in sensitivity, and 79.69% and 78.99% in specificity. The methods show an increasing performance with larger datasets, which provide us confidence on the validity of the proposed automation technique.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Ash Parameswaran
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

Detection and analysis of single event upsets in noisy digital imagers with small to medium pixels

Author: 
Date created: 
2020-09-02
Abstract: 

Camera sensors are shrinking, resulting in more defects seen through image analysis. Due to cosmic radiation, camera experience both permanent defects known as hot pixels and temporal defective spikes which are Single Event Upsets (SEUs). SEUs manifest themselves as temporal random bright areas in sequential dark-frame images that are taken with long exposure times. In the past, it was difficult to separate SEUs from noise in dark-frame images taken with DSLRs at high sensitivity levels (ISO) and cell phone cameras at modest sensitivity levels. However, recent software improvements in this research have enabled the analysis of defect rates in noisy digital imagers – by leveraging local area and pixel address distribution techniques. In addition, multiple experiments were performed to understand the relationship of SEUs and elevation. This study reports data from imagers with pixels ranging from 7 μm (DSLR cameras) down to 1.2 μm (cell phone cameras).

Document type: 
Thesis
File(s): 
Supervisor(s): 
Glenn Chapman
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

Far-field MEMS microphone array beamforming – Measurements, simulations, and design

Author: 
Date created: 
2020-08-21
Abstract: 

This thesis work determines the far-field array response patterns of micro-electromechanical system (MEMS) microphone arrays; and verifies these patterns employing experimental methods. Phase shifts and amplitude behaviour are simulated through finite element methods (FEM) using COMSOL Multiphysics, under both ideal and realistic conditions. Physical measurements are performed with microphone arrays using high accuracy audio analyzer equipment (Audio Precision APX555) to support and compare with mathematical and simulation conclusions. The effects of the packaging, mounting materials, and interference among elements on the array response patterns are studied using two-element microphone arrays. A new form of MEMS microphone array beamformer – a dynamic layout array beamformer – is introduced and simulated with the goals of improving flexibility, while lowering the complexity and power consumption, of MEMS microphone array systems. In addition to the acoustic signal recognition, a new approach is developed with a Xilinx Basys3 FPGA board to record and analyze the audio files using PmodMIC3 MEMS microphone devices. Applications based on the MEMS microphone array beamforming are introduced. Potential applications of the research to intelligent transportation system (ITS) moving vehicle direction of arrival (DOA) estimation are presented for further study.

Document type: 
Thesis
Supervisor(s): 
Bonnie Gray
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

Biomechanical analysis and simulation of backward falls with head impact in older adults

Author: 
Date created: 
2020-08-05
Abstract: 

This thesis examined the dynamics of backward falls in older adults involving head impact. Time-varying kinematics were extracted from digitizing videos of 11 real-life falls by residents of long-term care. The pelvis always impacted the ground before the head. On average, the head descended 1.2 m, and had a vertical velocity of 1.7 m/s just before it struck the ground. A novel dummy was used to examine how fall mechanics and compliant flooring affect head acceleration. Landing with a curved versus flat torso decreased peak rotational acceleration by 27% (4633 versus 5901 rad/s2). Landing with fixed versus freely rotating hips lowered peak translational accelerations by 36% (101.5 versus 158.7 g) and peak rotational accelerations by 38% (4168 versus 6366 rad/s2). The protective benefit of compliant flooring depended on torso curvature and hip stiffness. These results show that unexplored aspects of fall mechanics strongly influence head impact severity.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Stephen Robinovitch
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.

Deep learning-based multimedia content processing

Author: 
Date created: 
2020-07-24
Abstract: 

In the last few years, deep learning has revolutionized many applications in the field of multi-media content processing such as music information retrieval (MIR) and image compression, which are addressed in this thesis. In order to handle the challenges in acoustic-based MIR such as automatic music transcription, the video of the musical performances can be utilized. In Chapter 2, a new learning-based system for visually transcribing piano music using the convolutional neural networks and support vector machines is presented that achieves an average improvement of ~0.37 in terms of F1 score over the previous works. Another significant problem in MIR is music generation. In Chapter 3, a semi-recurrent hybrid model combining variational auto-encoder and generative adversarial network for sequential generation of piano music is introduced that achieves better results than previous methods. Auto-encoders have also been used as a perfect candidate for learned image compression, which has recently shown the potential to outperform standard codecs. Some efforts in integrating other computer vision tasks and image compression to improve the compression performance have also been made. In Chapter 4, a semantic segmentation-based layered image compression method is presented in which the segmentation map of the input is used in the compression procedure. Most learned image compression methods train multiple models for multiple bit rates, which increase the implementation complexity. In Chapter 5, we propose a variable-rate image compression model employing two novel loss functions and residual sub-networks in the auto-encoder. The proposed method outperforms the standard codecs and also previous learned variable-rate methods on Kodak image set. The state-of-the-art image compression has been achieved by utilizing joint hyper-prior and auto-regressive models. However, they suffer from the spatial redundancy of the low frequency information in the latents. In Chapter 6, we propose the first learned multi-frequency image compression approach that uses the recently developed octave convolutions to factorize the latents into high and low frequencies. As the low frequency is represented by a lower resolution, their spatial redundancy is reduced, which improves the compression rate. Our experiments show that the proposed scheme outperforms all standard codecs and learning-based methods in both PSNR and MS-SSIM metrics, and establishes the new state of the art for learned image compression on Kodak image set.

Document type: 
Thesis
File(s): 
Supervisor(s): 
Jie Liang
Department: 
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) Ph.D.