Objective: 3D scene reconstruction

Motivation: Novel view synthesis

Approach: Integrating multi-sensor fusion SLAM (LiDAR, Camera, IMU) with 3D Gaussian Splatting (3DGS)

Key Challenge: Addressing the modality gap between heterogeneous LiDAR and camera data

Outcome: Superior rendering accuracy compared to existing state-of-the-art methods

Objective: Autonomous navigation for robotic guide dogs across indoor/outdoor settings

Motivation: Enhancing mobility and independence for the visually impaired

Approach: Integrating LiDAR, Camera, and IMU sensors on a quadrupedal platform powered by an onboard embedded system

Key Challenge: Solving multimodal sensor fusion and avoiding dynamic obstacles in real-time

Outcome: Demonstrated stable autonomous locomotion and navigation in indoor testbeds

Objectives: Detecting arbitrary moving objects

Motivation: Improving safety in case that deep learning-based object detection fails or moving objects are not in traing datasets

Approach: 3D Computer Vision (Epipolar Geometry)

Key Challenge: Hard to tell if the objects seems to move because the vehicle is moving or they are moving by themselves

Outcome: NVIDIA DRIVE SDK

Objectives: Developing a robust multi-modal fusion SLAM system.

Motivation: Exploiting sensor redundancy and complementarity for enhanced reliability

Approach: Integration of LiDAR, Camera, IMU, and GPS data into a unified framework

Key Challenge: Handling non-uniform and asynchronous timestamps from disparate hardware

Outcome: Implementation of a continuous-time representation to effectively synchronize and fuse asynchronous measurements

Objective: Real-time RGB-D visual odometry

Motivation: Enabling seamless virtual object overlay for AR applications

Approach: Utilizing RGB-D sensor data for spatial mapping and tracking on AR hardware

Key Challenge: Robust outlier rejection for both photometric and geometric inconsistencies

Outcome: Stable localization performance under varying lighting conditions and sensor noise