KIMERA - Visual SLAM Perception

Status: Completed Tools: ROS, C++, KIMERA-VIO, OpenCV Lab: IvLabs

Abstract

This project implements visual SLAM (Simultaneous Localization and Mapping) using KIMERA, an open-source library for real-time metric-semantic visual-inertial odometry. KIMERA combines visual-inertial odometry with robust loop closure detection and 3D semantic mesh reconstruction to provide accurate localization and mapping in real-time.

System Overview

KIMERA-VIO (Visual-Inertial Odometry) processes stereo images and IMU data to estimate camera pose and build a 3D map. The pipeline consists of:

Front-end: Feature extraction and tracking using ORB features
IMU Preintegration: Fusing inertial measurements for robust pose estimation
Back-end: Sliding window optimization using factor graphs
Loop Closure: Place recognition using DBoW2 for global consistency

Key Features

Real-time performance on embedded platforms
Robust to lighting variations and motion blur
Semantic segmentation for object-aware mapping
Distributed mapping for large-scale environments

Results

The system was tested on the EuRoC MAV dataset achieving sub-centimeter localization accuracy. The semantic mesh reconstruction enabled obstacle avoidance for autonomous navigation.