Surveillance video data fusion

• Main Author: Wang, Simi
• Other Authors: Hunter, Gordon ; Ellis, Tim ; Orwell, James
• Published: Kingston University 2016
• Subjects: 006.3; Computer science and informatics
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.694082

The overall objective under consideration is the design of a system capable of automatic inference about events occurring in the scene under surveillance. Using established video processing techniques. low level inferences are relatively straightforward to establish as they only determine activities of some description. The challenge is to design a system that is capable of higher-level inference, that can be used to notify stakeholders about events having semantic importance. It is argued that re-identification of the entities present in the scene (such as vehicles and pedestrians) is an important intermediate objective, to support many of the types of higher level interference required. The input video can be processed in a number of ways to obtain estimates of the attributes of the objects and events in the scene. These attributes can then be analysed, or 'fused', to enable the high-level inference. One particular challenge is the management of the uncertainties, which are associated with the estimates, and hence with the overall inferences. Another challenge is obtaining accurate estimates of prior probabilities, which can have a significant impact on the final inferences. This thesis makes the following contributions. Firstly, a review of the nature of the uncertainties present in a visual surveillance system and quantification of the uncertainties associated with current techniques. Secondly, an investigation into the benefits of using a new high resolution dataset for the problem of pedestrain re-identification under various scenarios including occlusoon. This is done by combining state-of-art techniques with low level fusion techniques. Thirdly, a multi-class classification approach to solve the classification of vehicle manufacture logos. The approach uses the Fisher Discriminative classifier and decision fusion techniques to identify and classify logos into its correct categories. Fourthly, two probabilistic fusion frameworks were developed, using Bayesian and Evidential Dempster-Shafer methodologies, respectively, to allow inferences about multiple objectives and to reduce the uncertainty by combining multiple information sources. Fifthly, an evaluation framework was developed, based on the Kelly Betting Strategy, to effectively accommodate the additional information offered by the Dempster-Shafer approach, hence allowing comparisons with the single probabilistic output provided by a Bayesian analysis.