Mengzhen Chen
(Advisor: Prof. Dimitri Mavris]

will defend a doctoral thesis entitled,

Robust Autonomous Navigation Framework for Exploration in GPS-absent and Challenging Environment

On

Wednesday, April 10 at 2:00 p.m.
Collaborative Visualization Environment (CoVE)
Weber Space Science and Technology Building (SST II)

And

Microsoft Teams

Abstract
The benefits of autonomous systems have attracted the industry's attention during the past decade. Different kinds of autonomous systems have been applied to various fields such as transportation, agriculture, healthcare, etc. Tasks unable or risky to be completed by humans alone can now be handled by autonomous systems efficiently, and the labor cost has been greatly reduced. Among various kinds of tasks that an autonomous system can perform, the capability of an autonomous system to understand its surrounding environment is of great importance. Either using Unmanned Aircraft System (UAS) for package delivery or self-driving vehicles requires the autonomous system to be more robust during operation under different scenarios. This work will improve the robustness of autonomous systems under challenging and GPS-absent environments.

When exploring an unknown environment, if external information such as GPS signal is unavailable, mapping and localization are equally important and complementary. Therefore, simultaneously creating a map and localizing itself is essential. Under such conditions, Simultaneous Localization and Mapping (SLAM) was created in the robotics community to provide the capability of building a map for the surroundings of an autonomous system and localizing itself during operation. SLAM architecture has been designed for different kinds of sensors and scenarios during the past several decades. Among different SLAM categories, visual SLAM, which uses cameras as the sensors, outperforms others. It has the advantage of extracting rich information from images while other sensors alone are incapable. Since the images captured by the camera are treated as the inputs, therefore, the accuracy of the results will heavily depend on their quality. Most SLAM architecture can easily handle high-quality images or video streams, while poor-quality ones are still challenging. The first challenging scenario that the visual SLAM is facing is the motion blur scenario in which the performance of the visual SLAM will be severely downgraded. The other challenging scenario that the visual SLAM is facing is the low-light environment. Since the poor illumination condition has less information shared with the camera, it also downgrades the accuracy of the visual SLAM system. Furthermore, the visual SLAM adds an extra requirement for computational efficiency since the operation needs to be real-time.

Based on these observations, the research objective of this dissertation has been formed which is improving the visual SLAM performance under these two challenging conditions. In this dissertation, three research areas have been defined to achieve the overarching research objective. The first research area is focusing on developing the capabilities of recovering these poor-quality images captured under these challenging scenarios within real-time. Two highly efficient deep learning models, a single image deblurring model and low-light image enhancement model, have been developed and evaluated in this dissertation. The second research area is focusing on the uncertainty quantification for the results generated by the visual SLAM systems. Since some of the visual SLAM systems have nondeterministic behaviors, a statistical approach has been developed in this dissertation to reduce and factor out the uncertainties in the results and provides quantitative methods to performance evaluation. The third research area is focusing on creating a visual SLAM validation dataset that can be utilized for testing the performance under motion blur scenario since most of the existing dataset does not have enough blurriness or limited to indoor environment. In this dissertation, a synthetic blurry SLAM dataset has been created with the help of utilizing a physics-based virtual simulation environment. From a combination of the three research areas, a visual SLAM framework is proposed and tested with several visual SLAM datasets captured under the two challenging scenarios. Based on the experiment results, for the proposed visual SLAM framework, accuracy improvements have been observed through statistical approach for all the use cases when compared with the benchmark visual SLAM system. Therefore, the proposed visual SLAM framework in which the image enhancement modules have been added does improve the visual SLAM performance under challenging conditions.

Committee

·         Prof. Dimitri Mavris – School of Aerospace Engineering (Advisor)

·         Prof. Daniel P. Schrage – School of Aerospace Engineering

·         Prof. Kyriakos G. Vamvoudakis – School of Aerospace Engineering

·         Dr. Olivia J. Fischer – School of Aerospace Engineering

·         Dr. Youngjun Choi – United Parcel Service (UPS), Inc.