On-line scene understanding for closed loop control
This paper describes a rapid on-line system able to compute the semantics of outdoor scenes using dense stereo perception. Our main focus is to aid a robot to discover collision-free routes as an alternative to explore the environment during fall-back planning (localiser failure). The general scene...
| Main Authors: | , , , , |
|---|---|
| Format: | Conference item |
| Published: |
Robotics: Science and Systems
2016
|
| _version_ | 1797103550302519296 |
|---|---|
| author | Paz, LM Suleymanov, T Pinies, P Hester, G Newman, PM |
| author_facet | Paz, LM Suleymanov, T Pinies, P Hester, G Newman, PM |
| author_sort | Paz, LM |
| collection | OXFORD |
| description | This paper describes a rapid on-line system able to compute the semantics of outdoor scenes using dense stereo perception. Our main focus is to aid a robot to discover collision-free routes as an alternative to explore the environment during fall-back planning (localiser failure). The general scene understanding problem is formulated in a probabilistic framework that combines machine learning with continuous convex regularisation. In order to learn distinctive scene labels, our system relies on shallow classifiers in combination with a suite of contextual features derived from depth and colour cues. The proposed system is heterogeneous taking advantage of simultaneous GPGPU and multithreaded CPU to carry out important tasks such as dense depth map estimation, multi-labelling prediction and image segmentation. Extensive experiments on the KITTI dataset support the robustness of out system to derive collisionfree local routes. An accompanied video validates the system at live execution in an outdoor experiment with a wheeled robot exploring over hundreds of metres of trajectory. |
| first_indexed | 2024-03-07T06:21:41Z |
| format | Conference item |
| id | oxford-uuid:f2ebcd83-5a88-4597-9cb6-3b1d6cef1831 |
| institution | University of Oxford |
| last_indexed | 2024-03-07T06:21:41Z |
| publishDate | 2016 |
| publisher | Robotics: Science and Systems |
| record_format | dspace |
| spelling | oxford-uuid:f2ebcd83-5a88-4597-9cb6-3b1d6cef18312022-03-27T12:07:53ZOn-line scene understanding for closed loop controlConference itemhttp://purl.org/coar/resource_type/c_5794uuid:f2ebcd83-5a88-4597-9cb6-3b1d6cef1831Symplectic Elements at OxfordRobotics: Science and Systems2016Paz, LMSuleymanov, TPinies, PHester, GNewman, PMThis paper describes a rapid on-line system able to compute the semantics of outdoor scenes using dense stereo perception. Our main focus is to aid a robot to discover collision-free routes as an alternative to explore the environment during fall-back planning (localiser failure). The general scene understanding problem is formulated in a probabilistic framework that combines machine learning with continuous convex regularisation. In order to learn distinctive scene labels, our system relies on shallow classifiers in combination with a suite of contextual features derived from depth and colour cues. The proposed system is heterogeneous taking advantage of simultaneous GPGPU and multithreaded CPU to carry out important tasks such as dense depth map estimation, multi-labelling prediction and image segmentation. Extensive experiments on the KITTI dataset support the robustness of out system to derive collisionfree local routes. An accompanied video validates the system at live execution in an outdoor experiment with a wheeled robot exploring over hundreds of metres of trajectory. |
| spellingShingle | Paz, LM Suleymanov, T Pinies, P Hester, G Newman, PM On-line scene understanding for closed loop control |
| title | On-line scene understanding for closed loop control |
| title_full | On-line scene understanding for closed loop control |
| title_fullStr | On-line scene understanding for closed loop control |
| title_full_unstemmed | On-line scene understanding for closed loop control |
| title_short | On-line scene understanding for closed loop control |
| title_sort | on line scene understanding for closed loop control |
| work_keys_str_mv | AT pazlm onlinesceneunderstandingforclosedloopcontrol AT suleymanovt onlinesceneunderstandingforclosedloopcontrol AT piniesp onlinesceneunderstandingforclosedloopcontrol AT hesterg onlinesceneunderstandingforclosedloopcontrol AT newmanpm onlinesceneunderstandingforclosedloopcontrol |