Ming Jing Li

Visual Odometry Implementation

Implemented key components of a monocular visual odometry (VO) system in C++, focusing on algorithms for accurate pose estimation and 3D reconstruction from image sequences. Key components included image undistortion, perspective projection, Direct Linear Transformation (DLT) for the Perspective-n-Point (PnP) problem, Harris corner detection, KLT tracking, the normalized 8-point algorithm, and linear triangulation.

Nonparametric Estimation of Multiple Bridge Structures Using Vanishing Points and Outlier Handling

Developed a system for detecting multiple bridge spans and their orientations using vanishing points, which are essential for understanding the camera's perspective relative to structures. Estimated vanishing points from Hough lines to determine how the drone's camera is oriented concerning various bridge spans. Employed RANSAC algorithms to filter out outliers, significantly enhancing the accuracy of vanishing point estimations. Additionally, I implemented a method to identify distinct models representing different bridge spans and their corresponding vanishing points present in the scene.

Drone Tracking and Target Following Using Unscented Kalman Filter and Visual Servoing Techniques
In this project, I implemented an Unscented Kalman Filter (UKF) to estimate the motion of a target drone in a simulated environment, addressing nonlinear dynamics and measurement uncertainties. This state estimation feeds into the control system of a virtual UAV (follower). Utilizing Image-Based Visual Servoing (IBVS) techniques, the control system leverages visual feedback from YOLO-detected bounding boxes to adjust the follower UAV’s trajectory. The prototype focused on enabling the follower UAV to move forward in alignment with the target drone, demonstrating basic tracking functionality within the simulated environment.

Integration of GNSS and Visual-Inertial Odometry: Resolving Misalignment and Assessing GNSS Properties for Enhanced Localization

Resolved misalignment between the Visual-Inertial Odometry (VIO) frame and the global GNSS frame, crucial for accurate local-to-global coordinate anchoring. Additionally, monitored GNSS properties, including average pseudorange error and Doppler error standard deviations over time, to assess system performance. Analyzed the standard deviations of pseudorange errors, reflecting noise and variances in GNSS measurements, while evaluating Doppler errors critical for accurate velocity estimation. This comprehensive analysis provided valuable insights into system reliability and contributed to enhanced global localization accuracy.

Indoor Visual-Inertial Odometry Evaluation:

• Utilized an open-source VO evaluation tool to assess the performance of indoor VIO.
• Converted Optitrack motion captured ground truth data to ROS format for use in the evaluation process

Autonomous Drone Prototyping in Gazebo Simulator:

• Prototyped an autonomous drone in Gazebo with a D435i sensor, integrating open-source visual-inertial odometry and path planning algorithms, and interfaced with DroneKit and ArduPilot through SITL for flight controller emulation in a simulation environment.

Creating Dense Maps using Simulated Dual Omnidirectional Fisheye Cameras:

• Configured the geometry for two simulated omnidirectional fisheye cameras to align with the specifications of the open-source VINS-Fisheye algorithm.
• Replaced the existing Omnidirectional Fisheye Cameras models, as defined in VINS-Fisheye, with Gazebo's own omnidirectional fisheye camera model.

Omnidirectional Fisheye Visual-Inertial Localization on Orin Nano:

• Utilized open-source VINS-Fisheye to create a prototype for localization and dense mapping, using dual omnidirectional stereo cameras (IMX219) and a Pixhawk mini4 IMU.
• Identified challenges, such as those related to performance and the rolling shutter effect, and investigated solutions to enhance system performance and handle the rolling shutter effect.