BEST OUTCOME
From Scratch Implementation of Visual SLAM with Graph Optimization
Building and evaluating a full visual SLAM pipeline to understand real-world robot localization and mapping limitations.
DEMONSTRATED CAPABILITY
System-Level Perception Design
🎓 Graduate-level Systems Thinking
Designed and evaluated an interactive 3D reconstruction system by balancing neural rendering fidelity, geometric consistency, and hardware constraints to support scalable AR/VR experiences.
Analyzed and compared alternative reconstruction approaches to select methods that balance visual quality, computational efficiency, and interactivity under constrained resources.
Compared alternative technical approaches by analyzing trade-offs in accuracy, robustness, and resource constraints to inform system-level design decisions.
Compared alternative technical approaches by analyzing trade-offs in accuracy, robustness, and resource constraints to inform system-level design decisions.
What This Project Achieves
This project explores how visual SLAM systems estimate robot trajectories and maps by combining visual odometry with graph-based optimization. The team implemented a SLAM pipeline from scratch, applying feature-based visual odometry, loop closure detection, and pose graph optimization to both benchmark and self-collected datasets. Through experiments on the KITTI dataset and real-world camera data, the project highlights how optimization improves trajectory consistency while revealing practical constraints related to data quality, ground truth accuracy, and real-time performance.
How This Was Built — Key Highlights
This project implemented a full visual SLAM pipeline by integrating front-end visual odometry with back-end graph optimization. The approach combined feature tracking, pose estimation, loop closure detection, and optimization to evaluate SLAM performance across simulated, benchmark, and real-time settings.
Applied FAST feature detection with Lucas–Kanade optical flow to track keypoints efficiently between frames.
Estimated camera motion using essential matrix recovery and PnP, depending on data availability.
Implemented loop closure detection using feature matching and historical frame comparison.
Performed pose graph optimization with g2o using a sparse optimizer and Levenberg–Marquardt solver.
Evaluated the pipeline on the KITTI Odometry dataset, a small self-collected image set with ArUco markers, and a real-time RGB-D camera system.
Challenges
Developing and testing a visual SLAM system in realistic conditions introduced several technical and experimental challenges.
Sparse image frames and limited frame density reduced the reliability of visual odometry in self-collected datasets.
Inaccurate or noisy ground truth, particularly from printed ArUco markers, affected trajectory evaluation.
Feature tracking degraded under large rotations, motion blur, and low-texture environments.
Real-time operation required careful trade-offs between computational efficiency and pose estimation robustness.
Insights
Analysis across multiple datasets and demos revealed important insights into practical SLAM system design.
Graph-based optimization can reconcile odometry drift and observation noise, producing more consistent trajectories.
Visual odometry performance depends heavily on frame density, scene texture, and motion smoothness.
Combining lightweight front-end tracking with selective optimization supports real-time performance on CPU-based systems.
Reliable ground truth is critical for meaningful evaluation and debugging of SLAM pipelines.
Project Gallery
Academic Team Feedback
Feedback from the Project Lead—a researcher at MIT CSAIL and Co-Founder/CTO of a robotics company focused on manipulation and state estimation—highlighted the team’s strong grasp of core SLAM concepts and their ability to implement a complete system end to end. Drawing on his experience leading real-world robotics systems, he noted that the team successfully built and evaluated multiple SLAM variants across public benchmarks, self-collected datasets, and a real-time setup, demonstrating solid experimental depth. He specifically commended Atsuki’s clear explanation of graph-based SLAM and the group’s thorough experimentation using AR tags to validate system behavior. The Academic Coordinator additionally emphasized the team’s collaborative spirit and technical initiative, noting their effective use of the provided RGB camera, live on-site demonstration, and ability to deliver a working system under a highly compressed on-campus timeline.
Project Reflection
This project helped me build a solid foundation in computer vision techniques for robot localization and mapping. By implementing visual SLAM from scratch and testing it in both benchmark and real-world settings, I gained practical insight into how theoretical algorithms behave under real constraints and how optimization improves system reliability.