Localisation of ground range sensors using overhead imagery
<p>This thesis is about outdoor localisation using range sensors as an active sensor and cheap, publicly available satellite or `overhead' imagery as a prior map. Range sensors such as lidars and spinning FMCW radars are ideal for large-scale, outdoor autonomous navigation due to their lo...
| Үндсэн зохиолч: | |
|---|---|
| Формат: | Дипломын ажил |
| Хэл сонгох: | English |
| Хэвлэсэн: |
2023
|
| Нөхцлүүд: |
| _version_ | 1826312054825287680 |
|---|---|
| author | Tang, TY |
| author_facet | Tang, TY |
| author_sort | Tang, TY |
| collection | OXFORD |
| description | <p>This thesis is about outdoor localisation using range sensors as an active sensor and cheap, publicly available satellite or `overhead' imagery as a prior map. Range sensors such as lidars and spinning FMCW radars are ideal for large-scale, outdoor autonomous navigation due to their long sensing range, invariance to lighting conditions, and robustness against weather changes. Nevertheless, existing methods for range sensor localisation typically rely on prior maps collected from a previous mapping phase. On the other hand, off-the-shelf overhead imagery, such as public satellite images, is readily available almost anywhere in the world and can be acquired easily from the internet with little cost or effort.</p>
<p>Public overhead imagery can capture geometric cues of the scene also observable by ground lidars and radars, therefore having the capability to act as a map for range sensor localisation. In particular, in corner case scenarios where the prior sensory map is unusable or unavailable, for example if the robot travels to a place it has not visited before, public overhead imagery can act as an alternative map source for range sensor localisation as a fall-back choice. Under normal operation conditions, the localisation result by comparing range sensor data against overhead imagery can act as an additional information source for redundancy.</p>
<p>In this thesis, we present various methods to solve the localisation of a ground range sensor using overhead imagery by learning from data, enabling them to adapt to different environments. This surpasses the methods in literature which employ hand-crafted features designed for only specific types of scenery. Specifically, we address both topological localisation, also known as place recognition, and metric localisation in overhead imagery maps. Furthermore, we investigate self-supervised strategies that allow the tasks to be learned without accurate ground truth data.</p> |
| first_indexed | 2024-03-07T08:21:55Z |
| format | Thesis |
| id | oxford-uuid:96d7988f-703f-4057-9355-0acbf9591dc8 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T08:21:55Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | oxford-uuid:96d7988f-703f-4057-9355-0acbf9591dc82024-01-31T12:30:06ZLocalisation of ground range sensors using overhead imageryThesishttp://purl.org/coar/resource_type/c_db06uuid:96d7988f-703f-4057-9355-0acbf9591dc8RoboticsEnglishHyrax Deposit2023Tang, TY<p>This thesis is about outdoor localisation using range sensors as an active sensor and cheap, publicly available satellite or `overhead' imagery as a prior map. Range sensors such as lidars and spinning FMCW radars are ideal for large-scale, outdoor autonomous navigation due to their long sensing range, invariance to lighting conditions, and robustness against weather changes. Nevertheless, existing methods for range sensor localisation typically rely on prior maps collected from a previous mapping phase. On the other hand, off-the-shelf overhead imagery, such as public satellite images, is readily available almost anywhere in the world and can be acquired easily from the internet with little cost or effort.</p> <p>Public overhead imagery can capture geometric cues of the scene also observable by ground lidars and radars, therefore having the capability to act as a map for range sensor localisation. In particular, in corner case scenarios where the prior sensory map is unusable or unavailable, for example if the robot travels to a place it has not visited before, public overhead imagery can act as an alternative map source for range sensor localisation as a fall-back choice. Under normal operation conditions, the localisation result by comparing range sensor data against overhead imagery can act as an additional information source for redundancy.</p> <p>In this thesis, we present various methods to solve the localisation of a ground range sensor using overhead imagery by learning from data, enabling them to adapt to different environments. This surpasses the methods in literature which employ hand-crafted features designed for only specific types of scenery. Specifically, we address both topological localisation, also known as place recognition, and metric localisation in overhead imagery maps. Furthermore, we investigate self-supervised strategies that allow the tasks to be learned without accurate ground truth data.</p> |
| spellingShingle | Robotics Tang, TY Localisation of ground range sensors using overhead imagery |
| title | Localisation of ground range sensors using overhead imagery |
| title_full | Localisation of ground range sensors using overhead imagery |
| title_fullStr | Localisation of ground range sensors using overhead imagery |
| title_full_unstemmed | Localisation of ground range sensors using overhead imagery |
| title_short | Localisation of ground range sensors using overhead imagery |
| title_sort | localisation of ground range sensors using overhead imagery |
| topic | Robotics |
| work_keys_str_mv | AT tangty localisationofgroundrangesensorsusingoverheadimagery |