Illumination Invariant Outdoor Perception
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Ramakrishnan, RishiAbstract
This thesis proposes the use of a multi-modal sensor approach to achieve illumination invariance in images taken in outdoor environments. The approach is automatic in that it does not require user input for initialisation, and is not reliant on the input of atmospheric radiative ...
See moreThis thesis proposes the use of a multi-modal sensor approach to achieve illumination invariance in images taken in outdoor environments. The approach is automatic in that it does not require user input for initialisation, and is not reliant on the input of atmospheric radiative transfer models. While it is common to use pixel colour and intensity as features in high level vision algorithms, their performance is severely limited by the uncontrolled lighting and complex geometric structure of outdoor scenes. The appearance of a material is dependent on the incident illumination, which can vary due to spatial and temporal factors. This variability causes identical materials to appear differently depending on their location. Illumination invariant representations of the scene can potentially improve the performance of high level vision algorithms as they allow discrimination between pixels to occur based on the underlying material characteristics. The proposed approach to obtaining illumination invariance utilises fused image and geometric data. An approximation of the outdoor illumination is used to derive per-pixel scaling factors. This has the effect of relighting the entire scene using a single illuminant that is common in terms of colour and intensity for all pixels. The approach is extended to radiometric normalisation and the multi-image scenario, meaning that the resultant dataset is both spatially and temporally illumination invariant. The proposed illumination invariance approach is evaluated on several datasets and shows that spatial and temporal invariance can be achieved without loss of spectral dimensionality. The system requires very few tuning parameters, meaning that expert knowledge is not required in order for its operation. This has potential implications for robotics and remote sensing applications where perception systems play an integral role in developing a rich understanding of the scene.
See less
See moreThis thesis proposes the use of a multi-modal sensor approach to achieve illumination invariance in images taken in outdoor environments. The approach is automatic in that it does not require user input for initialisation, and is not reliant on the input of atmospheric radiative transfer models. While it is common to use pixel colour and intensity as features in high level vision algorithms, their performance is severely limited by the uncontrolled lighting and complex geometric structure of outdoor scenes. The appearance of a material is dependent on the incident illumination, which can vary due to spatial and temporal factors. This variability causes identical materials to appear differently depending on their location. Illumination invariant representations of the scene can potentially improve the performance of high level vision algorithms as they allow discrimination between pixels to occur based on the underlying material characteristics. The proposed approach to obtaining illumination invariance utilises fused image and geometric data. An approximation of the outdoor illumination is used to derive per-pixel scaling factors. This has the effect of relighting the entire scene using a single illuminant that is common in terms of colour and intensity for all pixels. The approach is extended to radiometric normalisation and the multi-image scenario, meaning that the resultant dataset is both spatially and temporally illumination invariant. The proposed illumination invariance approach is evaluated on several datasets and shows that spatial and temporal invariance can be achieved without loss of spectral dimensionality. The system requires very few tuning parameters, meaning that expert knowledge is not required in order for its operation. This has potential implications for robotics and remote sensing applications where perception systems play an integral role in developing a rich understanding of the scene.
See less
Date
2015-09-24Faculty/School
Faculty of Engineering and Information Technologies, School of Aerospace, Mechanical and Mechatronic EngineeringAwarding institution
The University of SydneyShare