Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.606133
Title: Semantic labelling of road scenes using supervised and unsupervised machine learning with lidar-stereo sensor fusion
Author: Osgood, Thomas J.
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2013
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
At the highest level the aim of this thesis is to review and develop reliable and efficient algorithms for classifying road scenery primarily using vision based technology mounted on vehicles. The purpose of this technology is to enhance vehicle safety systems in order to prevent accidents which cause injuries to drivers and pedestrians. This thesis uses LIDAR–stereo sensor fusion to analyse the scene in the path of the vehicle and apply semantic labels to the different content types within the images. It details every step of the process from raw sensor data to automatically labelled images. At each stage of the process currently used methods are investigated and evaluated. In cases where existingmethods do not produce satisfactory results improvedmethods have been suggested. In particular, this thesis presents a novel, automated,method for aligning LIDAR data to the stereo camera frame without the need for specialised alignment grids. For image segmentation a hybrid approach is presented, combining the strengths of both edge detection and mean-shift segmentation. For texture analysis the presented method uses GLCM metrics which allows texture information to be captured and summarised using only four feature descriptors compared to the 100’s produced by SURF descriptors. In addition to texture descriptors, the ìD information provided by the stereo system is also exploited. The segmented point cloud is used to determine orientation and curvature using polynomial surface fitting, a technique not yet applied to this application. Regarding classification methods a comprehensive study was carried out comparing the performance of the SVM and neural network algorithms for this particular application. The outcome shows that for this particular set of learning features the SVM classifiers offer slightly better performance in the context of image and depth based classification which was not made clear in existing literature. Finally a novel method of making unsupervised classifications is presented. Segments are automatically grouped into sub-classes which can then be mapped to more expressive super-classes as needed. Although the method in its current state does not yet match the performance of supervised methods it does produce usable classification results without the need for any training data. In addition, the method can be used to automatically sub-class classes with significant inter-class variation into more specialised groups prior to being used as training targets in a supervised method.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.606133  DOI: Not available
Keywords: HE Transportation and Communications ; QA76 Electronic computers. Computer science. Computer software ; TK Electrical engineering. Electronics Nuclear engineering
Share: