Visual-Inertial SLAM via EKF

The Simultaneous Localization and Mapping (SLAM) problem has been one of the most popular research areas from its coinage. With the breakthrough of robotics and the usage of many related smart devices and observation sensors, the problem of accurately locating the device and building a realtime map of its surrounding environment becomes a popular subject with dense literature in finding a way to best utilize all the collected data and recover the trajectory and the map. Among many proposed solutions, visual-inertial SLAM is one approach that takes advantage of IMU and stereo camera measurements to robustly perform this task. In this project, we discuss how to implement visual-inertial SLAM on a moving vehicle equipped with the camera and IMU.

This project implements visual-inertial simultaneous localization and mapping (SLAM) using an extended Kalman filter (EKF). Using synchronized measurements from an inertial measurement unit (IMU) and visual landmark features extracted from a stereo camera provided with its intrinsic and extrinsic calibration equipped on a vehicle, I estimated its trajectory and reconstructed the landmarks of its surrounding environment via two approaches. I first solved this problem by separating out the localization and visual mapping tasks and solving them individually to obtain a dead reckoning result, then compared this result with EKF approach which considers the correlation between the landmarks and the vehicle.

Project Report

• Gaopo Huang. "Visual-inertial SLAM via Extended Kalman Filter". UCSD, student paper

Visualization

• Implemented SLAM output for vechicle trajectory and feature mapping.

File structure

├── code
│ ├── training
│ │ ├── main.py: #Visual inertial SLAM implementation
│ │ ├── pr3_util.py: #Helper functions for graphing and pose kinematics
├── data #Contains all the train data
├── dataset03figs
│ │ ├── LocalizationOnly #part(a)
│ │ ├── MapOnly_W1E-02_V_02_lm_05 #part(b)
│ │ ├── SLAM_W1E-02_V_02_lm_05 #part(c) SLAM only
│ │ ├── Comparison_W1E-02_V_02_lm_05 #part(c) SLAM vs trajectory
│ │ └── ...
├── dataset10figs
│ │ ├── LocalizationOnly #part(a)
│ │ ├── MapOnly_W1E-02_V_02_lm_05 #part(b)
│ │ ├── SLAM_W1E-02_V_02_lm_05 #part(c) SLAM only
│ │ ├── Comparison_W1E-02_V_02_lm_05 #part(c) SLAM vs trajectory
│ │ └── ...
├── imgs
├── reports
│── README.md

Usage

main.py provides the basic visual-inertial SLAM implementation. It can run the dead reckoning prediction and update steps and also visual-inertial SLAM combined. Some parameters that can be set in the main file before running:

• full_evaluate: test different noise tuning.
• live_plot_update: show a live updating plot
• subsample_rate: subsampling the feature files
• save_freq: save figure frequency
• All noise tuning including W, V, and landmark covariance initialization noise