ST-VIO :Online Calibration of a Single-Track Ground Vehicle Dynamics Model by Tight Fusion with Visual-Inertial Odometry
Haolong Li1 and Joerg Stueckler1
ST-VIO (Single-Track Visual Inertial Odometry) is a novel approach designed to enhance the navigation and motion estimation capabilities of wheeled mobile robots. This system tightly integrates a single-track dynamics model with visual inertial odometry (VIO) to improve the accuracy of motion predictions and control actions for navigation planning.
The Single Track Dynamics Model, also known as the Bicycle Model or the Rigid Rider Model, is a widely used model in the field of vehicle dynamics for analyzing and predicting the motion and stability of two-wheeled vehicles, such as motorcycles, bicycles, and other similar vehicles. The model simplifies the vehicle by representing it as a single rigid body with two wheels, one in the front and one in the rear. The front wheel is assumed to be steerable, while the rear wheel is fixed. The model assumes that the vehicle moves in a plane, neglecting any vertical motion or roll dynamics.
Visual_Inertial_Odometry.mp4
Visual Inertial Odometry (VIO) is a technique used in robotics and autonomous systems for estimating the pose (position and orientation) of a moving platform, such as a robot, vehicle, or drone, by fusing data from visual sensors (e.g., cameras) and inertial sensors (e.g., accelerometers and gyroscopes).Here's how VIO works:
- Visual sensors (cameras) are used to track features (e.g., corners, edges) in the environment across consecutive frames.
- By analyzing the motion of these features, the system can estimate the relative motion of the platform between frames, providing incremental pose updates.
- However, VO is prone to drift over time due to accumulated errors and is sensitive to rapid motions or lack of distinguishable features.
- Inertial sensors (accelerometers and gyroscopes) provide linear acceleration and angular velocity measurements, respectively.
- These measurements can be integrated to estimate the platform's pose, but they suffer from unbounded drift due to sensor biases and integration of noise over time.
- VIO combines the strengths of both visual and inertial sensors by fusing their data in a tightly coupled manner.
- The visual measurements provide absolute pose updates to correct the drift in the inertial measurements, while the inertial measurements help maintain pose estimates during rapid motions or when visual features are unavailable.
- Imagine you have a robot with a camera and an inertial measurement unit (IMU) mounted on it. The camera and IMU have a fixed position and orientation relative to the robot's body frame (called the extrinsic transformation).
- The visual-inertial odometry (VIO) system estimates the 6-DoF pose (position and orientation) of the camera-IMU frame relative to the world frame at each timestep, using the camera images and IMU measurements.
- We are interested in the 3-DoF motion (2D position and heading angle) of the robot's body frame relative to the world frame, as the robot is a ground vehicle constrained to move on a flat surface. To compare the 6-DoF camera-IMU motion estimated by the VIO with the 3-DoF robot body motion predicted by the dynamics model, we need to transform the VIO estimate into the robot's body frame.
A 1/10 scale electric car
equipped with a Realsense T265 stereo-fisheye camera with
built-in IMU.
- Evaluated tracking and prediction accuracy in various scenarios and compare with the original VIO.
The tracking accuracy of the proposed ST-VIO method is evaluated by comparing the relative pose error (RPE) with the original VIO approach. The results show that ST-VIO generally improves trajectory accuracy for indoor sequences. However, the outdoor data with bumpy terrain poses challenges for the single-track dynamics model.