Optimal estimation algorithms play a critical role in the area of Unmanned Ground Vehicle&nbsp;(UGV) outdoor localization. Outdoor navigation algorithms that provide autonomy are required in a variety of industries. Effective outdoor navigation requires a reliable and robust outdoor localization algorithm. However, due to obstructions such as buildings and trees, the Global Navigation Satellite System (GNSS) becomes inaccurate or inaccessible in some cases. Additionally, odometers have their own drawbacks. They are susceptible to error due to high slippage in off-road terrains, misalignment of right and left wheels, and uneven tire pressures. With the Inertial Measurement Unit (IMU), the accelerometer values cannot be used as error accumulates over time even after few centimeters of movement.Traditionally, Extended Kalman Filters (EKF) are used in the following setup GNSS/Odometer, GNSS/IMU, GNSS/Optical flow, and GNSS/IMU/Odometer. All of the above combinations have their own advantages but with accuracy limitations in a multi-terrain environment. To address these limitations, we introduce a neural network-assisted GNSS Real Time Kinematics (RTK), IMU, odometer, and ground-facing camera with an optical flow. An EKF is formulated to combine the data from these sensors along with terrain-specific covariance matrices to improve localization.The covariance matrix has a major impact on positioning accuracy in the EKF localization setup, and it can also cause the filter to diverge when the GNSS read is offline for a long period of time. Despite numerous studies on covariance estimation, the ability of conventional methods to adapt to a wide range of environments remains limited.In one of our previous works, EKF was utilized to predict the vehicle pose using an odometer, GNSS-RTK, and IMU. The results were compared using camera verification from a quadcopter. In another previous work by the author, the accuracy and reliability of GNSS, IMU, and speedometer have been explored but due to GNSS long-term dropout error have been accumulated.Currently, we propose the use of GNSS-RTK/IMU/odometer/optical flow-based EKF system to improve state estimation and compare it with GNSS/IMU/odometer/optical flow. The ground-facing camera will be integrated with image processing techniques such as optical flow to get linear and angular velocities. Additionally, a convolutional neural network (CNN) will be used to detect the terrain conditions and update the system covariance matrices accordingly. This will be done by first training the CNN on different terrain conditions with terrain-specific covariance matrices. Later on, the trained CNN will predict the probability of using a specific covariance matrix for the next EKF pose update step.The system predicted pose will be validated with an acceptable error of +-2 centimeters. The first test is on a 4 meter by 4 meter grid area quadcopter camera for validation. The second test is on a 10 meter by 10 meter grid area with a section-based quadcopter camera validation. The quadcopter will fly over the first region (4x4) and once the UGV goes outside the camera frame, the quadcopter will move to the next location to maintain a high level of accuracy as compared to flying the quadcopter at a higher altitude. The third test will be on a curved trajectory. These three tests encompass the different modes of trajectory for the UGV, which are straight line, and skid steering.

UGV Localization With AI-Assisted EKF for Multi-Terrain Environments

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.