Introduction

The AutoRally platform stands as an advanced testbed designed for cutting-edge research in self-driving vehicles. AutoRally provides an opportunity for researchers and enthusiasts to explore aggressive autonomous off-road driving. This project is my capstone project done under Dr. Tsiotras' guidance at Georgia Tech's DCSL lab.

The core of the project revolves around implementing a localization system for the vehicle that relies on cameras and IMU. Traditional global navigation satellite systems (GNSS) based methods suffer from signal interference in noisy signal environments. Also, they suffer from signal blockage in indoor or other GNSS denied areas which makes them unreliable in such environments.

Project Report Project Poster Github Repository

Goals

The main objective of the project is the implementation of a reliable Visual Inertial Odometry (VIO) system on the AutoRally platform used for autonomous vehicle research. This involves research into the state of the art VIO systems, evaluating the ones that would be best suited for the AutoRally platform in simulation, interfacing with the hardware to get these algorithms on the track and finally investigating and optimizing their performance as per the available hardware.

1. Review State of the Art VIO systems
2. Calibrate them for Autorally Platform
3. Finetune the systems for optimal performance

Work Done

In the first semester, literature review of Visual-Inertial Odometry (VIO) was done. This involved a deep dive into feature tracking, filtering and particle based estimators, and camera calibration. Autorally has two stereo cameras and an IMU for VIO.

Four VIO systems
1. XIVO
2. ORBslam3
3. OpenVINS
4. Rovio
were carefully chosen based on various factor and subsequent work started on calibrating hardware for these systems.

Utilizing Allan Variance ROS, essential IMU parameters like bias and noise were precisely estimated, enhancing the accuracy of the inertial measurements. For a robust representation of the camera's behavior, the team adopted the Pinhole Radial Tangential camera model. Kalibr played a pivotal role in calibrating camera projection and distortion coefficients, ensuring seamless synchronization with the IMU stream. All tools used had ROS wrappers for ease of use.

Initial calibration for IMU was done using Allan Variance deviation using an online tool. This gave essential plots of the Allan deviations. OpenCV and Kalibr were used to estimate camera distortion and projection matrix coefficients. The reprojection errors and undistorted images are shown here.

Results

Initial testing was done in an outdoor environment. Most packages performed poorly. Some reasons for poor performance were:
1. Noisy compressed image data in grassy environments
2. Sudden variation in exposure when directly facing the sun
3. Tracking undesired features like on vehicle chassis

Lot of parameter fine-tuning was done to optimize the results. This included varying the feature tracker, number of features tracked, features to use in the estimation step, etc.

OpenVINS gave the most promising results and the results are shown here. The results weren't the best especially with lots of error accumulation. In order to simplify the problem, an indoor environment was next used to validate the performance.

GNSS localization is a major issue in indoor environments which is why validation was not possible with the exact trajectory. So the verification was done by having the vehicle follow fixed loops on an indoor track.
1. 3 loops on an oblong track
2. 2 loops in another shorter oblong track
3. 5 loops on a circular track

Since the control was manual, the loops were not perfectly followed. However, as can be seen from feature tracking data, a consistent indoor environment helps in reliably tracking features for state estimation.

Overall, the results are much more accurate as compared to before and it can be shown that the algorithm works reliably on the indoor platform. In future, the results can be made even better with further loop closures.

Introduction

Goals

Work Done

Results

Contact me!

Linkedin

Github