- Title
- Towards Bayesian optical flow
- Creator
- Farnworth, Timothy
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2023
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Autonomous systems play a crucial role in various aspects of modern transportation, agriculture, and healthcare. Navigation is a critical component of autonomous systems, enabling them to understand their position and interact with the surrounding environment. This requires the use of different sensing technologies such as GPS, IMU, and vision. Among these, vision sensors are exceedingly versatile and rich in information. One of the primary channels of visual information used in nature is observed motion in the scene, known as optical flow. With a brain weighing less than a milligram, bees exploit optical flow to successfully navigate to and from food sources that may be located 10km away from their hive. Considering its tiny brain, researchers regard the bee as the minimal working example of a complete visual navigation solution. Decades of research have sought to replicate the bee’s remarkable navigation capabilities in engineered solutions. In practice, optical flow is measured by applying algorithms to image sequences captured by cameras and has been leveraged in computer science and engineering to improve position and orientation (pose) estimates of navigation systems. However, the variability of optical flow measurements is difficult to characterise, as it depends not only on the imaging sensor but also on environmental lighting and texture, as well as the algorithms employed. Previous models have assumed a Gaussian noise distribution, which is convenient for posing egomotion estimation as a Gaussian least-squares problem. However, recent research has shown that this assumption can be inadequate, leading to errors in pose estimation and uncertainty measures. Consequently, there are ongoing efforts to develop new models and algorithms that improve the estimates of pose and motion in autonomous systems that use vision-based navigation. This thesis aims to address these challenges. Firstly, a novel texture-based likelihood model for optical flow is developed that utilises a Laplace-Cauchy mixture (LCM) distribution. The likelihood adheres to peaks and tails of the optical flow error distributions observed in empirical data, whereas the traditional Gaussian assumption is restrictive in describing this phenomenon. Compared to the empirical characteristics, the LCM distribution has 94% less error than the traditional Gaussian model and 48% less error than contemporary models, according to the Kolmogorov- Smirnov statistic. It is demonstrated on the KITTI dataset to improve visual odometry estimation accuracy and minimise position drift by at least 39%, surpassing the performance of previous models. Secondly, to use the novel LCM likelihood model within a Gaussian state estimator, we present VIVO, an approximate Bayes visual odometry solution based on variational inference. VIVO fuses information from a Wiener motion prior and optical flow measurements via the LCM likelihood model. This leads to a 43% reduction in drift compared to maximum likelihood approaches with the same likelihood model. Results show that VIVO results in a 57%–71% reduction in estimator position drift compared to contemporary methods. The presented work ushers the field of robotics toward a Bayesian treatment of optical flow measurements enabling the fusion of flexible non-Gaussian optical flow error models within a Gaussian assumed density filter. The development of the LCM likelihood model reduces the chasm between approximate models and empirical data, enabling justifiable reasoning with optical flow measurements. Overall, the work presented in this thesis enhances the componentry required to perform Bayesian sensor fusion with vision data, enabling the inclusion of vision with sensors such as GNSS, LiDAR and IMU, strengthening trusted autonomy.
- Subject
- Bayesian optical flow; assumed density filtering; localisation; sensor fusion
- Identifier
- http://hdl.handle.net/1959.13/1487824
- Identifier
- uon:52263
- Rights
- Copyright 2023 Timothy Farnworth
- Language
- eng
- Full Text
- Hits: 1113
- Visitors: 1167
- Downloads: 102
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | ATTACHMENT01 | Thesis | 32 MB | Adobe Acrobat PDF | View Details Download | ||
| View Details Download | ATTACHMENT02 | Abstract | 453 KB | Adobe Acrobat PDF | View Details Download |