The research reported in this thesis has examined means of enhancing the performance of the Optimal Trade-off Maximum Average Correlation Height (OT-MACH) filter for target detection in Forward Looking Infra-Red (FLIR) imagery acquired from a helicopter and border security FLIR camera in northern Kuwait. The data acquired with these FLIR sensors allows real-world evaluation of the comparative performance of the various filters that have been developed in the thesis. The results obtained have been quantified using well known performance measures such as Peak to Side-lobe Ratio (PSR) and Total Detection Error (TDE). The initial focus was to study the effect of modifying the OT-MACH parameters on the correlation metrics. A new optimisation technique has been presented, which computes statistically the filter alpha parameter associated with controlling the response of the filter to clutter noise. A further modification of the OT-MACH filter performance using the Difference of Gaussian bandpass filter (named the D-MACH filter) as a pre-processing stage has been described. The D-MACH has been applied to several test images containing single and multiple targets in the scene. Enhanced performance of the modified filter is demonstrated with improved metrics being obtained with less false side peaks in the correlation plane, especially when multiple targets are present in the test images. A further pre-processing technique was investigated using the Rayleigh distribution as a pre-processing filter (named the R-MACH filter). The R-MACH filter has been applied to multiple target types with tests conducted across various image data sets. The filter demonstrated an improvement over the Difference of Gaussian filter in terms of 6 reducing the number of parameters needing to be tuned whilst producing further enhanced correlation plane metrics. Finally, recommendations for future work has been made to improve the use of the OT-MACH filter in target detection and identification. A novel training image representation is proposed for further investigation, which will minimise the computational intensity of using the MACH filter for unconstrained object recognition.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:629045 |
| Date | January 2014 |
| Creators | Alkandri, Ahmad |
| Publisher | University of Sussex |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://sro.sussex.ac.uk/id/eprint/49954/ |
Page generated in 0.0015 seconds