Computationally-efficient visual inertial odometry for autonomous vehicle

This thesis presents the design, implementation, and validation of a novel nonlinearfiltering based Visual Inertial Odometry (VIO) framework for robotic navigation in GPSdenied environments. The system attempts to track the vehicle’s ego-motion at each time instant while capturing the benefits of both the camera information and the Inertial Measurement Unit (IMU). VIO demands considerable computational resources and processing time, and this makes the hardware implementation quite challenging for micro- and nanorobotic systems. In many cases, the VIO process selects a small subset of tracked features to reduce the computational cost. VIO estimation also suffers from the inevitable accumulation of error. This limitation makes the estimation gradually diverge and even fail to track the vehicle trajectory over long-term operation. Deploying optimization for the entire trajectory helps to minimize the accumulative errors, but increases the computational cost significantly. The VIO hardware implementation can utilize a more powerful processor and specialized hardware computing platforms, such as Field Programmable Gate Arrays, Graphics Processing Units and Application-Specific Integrated Circuits, to accelerate the execution. However, the computation still needs to perform identical computational steps with similar complexity. Processing data at a higher frequency increases energy consumption significantly. The development of advanced hardware systems is also expensive and time-consuming. Consequently, the approach of developing an efficient algorithm will be beneficial with or without hardware acceleration. The research described in this thesis proposes multiple solutions to accelerate the visual inertial odometry computation while maintaining a comparative estimation accuracy over long-term operation among state-ofthe- art algorithms. This research has resulted in three significant contributions. First, this research involved the design and validation of a novel nonlinear filtering sensor-fusion algorithm using trifocal tensor geometry and a cubature Kalman filter. The combination has handled the system nonlinearity effectively, while reducing the computational cost and system complexity significantly. Second, this research develops two solutions to address the error accumulation issue. For standalone self-localization projects, the first solution applies a local optimization procedure for the measurement update, which performs multiple corrections on a single measurement to optimize the latest filter state and covariance. For larger navigation projects, the second solution integrates VIO with additional pseudo-ranging measurements between the vehicle and multiple beacons in order to bound the accumulative errors. Third, this research develops a novel parallel-processing VIO algorithm to speed up the execution using a multi-core CPU. This allows the distribution of the filtering computation on each core to process and optimize each feature measurement update independently. The performance of the proposed visual inertial odometry framework is evaluated using publicly-available self-localization datasets, for comparison with some other open-source algorithms. The results illustrate that a proposed VIO framework is able to improve the VIO’s computational efficiency without the installation of specialized hardware computing platforms and advanced software libraries.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.