AUV Localization
For my senior year capstone project, I worked with a team (three other members in the Fall, with two new members in the Spring) to develop the localization for a custom Autonomous Underwater Vehicle in a GPS denied environment (ie. underwater).
In the previous year, the team focused on assembling the AUV, and developing the controls. However they ran into difficulties while testing the control algorithms because the localization was extremely unreliable. Previously, there was no official localization algorithm, and instead the team simply integrated accelerometer data fused with gyroscope data. Below is an illustration of the state of the localization prior to my team's work.
Thus our objective was to improve the localization of the AUV. The AUV was equipped with an IMU (accelerometer, gyroscope, and magnetometer), pressure sensor, and GPS. Typically, this sensor suite would be sufficient, say for a quadcopter, but underwater localization is particularly difficult because there is no GPS. Without GPS, there is no absolute positional measurements, and thus localization based on inertial measurements is unusable due to drift (as was the case in the above video).
Thus the project was divided into two main components: Developing a system to obtain absolute positioning, and developing the actual localization algorithm.
To get absolute positioning, the team decided to use acoustic localization, specifically One-Way Travel-Time Inverted Ultra-Short Baseline Localization (see paper for details ). With a microphone array on the AUV and a single source beacon at the surface of the water, the AUV can be localized relative to the beacon based on a time of flight for each microphone.
As for the localization algorithm, we decided to develop an Extended Kalman Filter (EKF). The EKF is an industry standard algorithm for fusing global and inertial data along with a motion model.
Ultimately, we were unable to unite the two aspects of the project due to time constraints. Specifically, the acoustic localization system was setup, but we encountered signal attenuation issues that prevented the Matched Filter from reliably determining the time of flight of the signals. On the other hand, the 2D EKF was fully developed and tested with simulated data. A dynamic motion model was experimentally determined by a linear and angular step test. A SimuLink model was used, and sample data was generated with Additive White Gaussian Noise. Then the EKF (implemented in MATLAB) was tested, and compared to the ground truth data.
Below is a comparison of the 2D EKF and direct integration (the localization method of last year's team) to show the improvement.
Below is a closer look at the performance of the 2D EKF.
Ultimately the 2D EKF performed very well, with the estimated location always staying within one sigma of the true location. The sampled data was generated using the parameter specifications of the sensors (IMU, and Acoustic Localization) to be as realistic as possible. The algorithm was also tested using non-zero biases in the noise and still performed well.
There have been many details left out, so for more information see the technical report, or feel free to reach out to me with questions.