Masters Degrees (Computer Engineering)

Permanent URI for this collectionhttps://hdl.handle.net/10413/6913

Browse

Now showing 1 - 3 of 3

Gaussian mixture model classifiers for detection and tracking in UAV video streams.
(2017) Pillay, Treshan.; Naidoo, Bashan.
Manual visual surveillance systems are subject to a high degree of human-error and operator fatigue. The automation of such systems often employs detectors, trackers and classifiers as fundamental building blocks. Detection, tracking and classification are especially useful and challenging in Unmanned Aerial Vehicle (UAV) based surveillance systems. Previous solutions have addressed challenges via complex classification methods. This dissertation proposes less complex Gaussian Mixture Model (GMM) based classifiers that can simplify the process; where data is represented as a reduced set of model parameters, and classification is performed in the low dimensionality parameter-space. The specification and adoption of GMM based classifiers on the UAV visual tracking feature space formed the principal contribution of the work. This methodology can be generalised to other feature spaces. This dissertation presents two main contributions in the form of submissions to ISI accredited journals. In the first paper, objectives are demonstrated with a vehicle detector incorporating a two stage GMM classifier, applied to a single feature space, namely Histogram of Oriented Gradients (HoG). While the second paper demonstrates objectives with a vehicle tracker using colour histograms (in RGB and HSV), with Gaussian Mixture Model (GMM) classifiers and a Kalman filter. The proposed works are comparable to related works with testing performed on benchmark datasets. In the tracking domain for such platforms, tracking alone is insufficient. Adaptive detection and classification can assist in search space reduction, building of knowledge priors and improved target representations. Results show that the proposed approach improves performance and robustness. Findings also indicate potential further enhancements such as a multi-mode tracker with global and local tracking based on a combination of both papers.
Parallel patch-based volumetric reconstruction from images.
(2014) Jermy, Robert Sydney.; Naidoo, Bashan.; Tapamo, Jules-Raymond.
Three Dimensional (3D) reconstruction relates to the creating of 3D computer models from sets of Two Dimensional (2D) images. 3D reconstruction algorithms tend to have long execution times, meaning they are ill suited to real time 3D reconstruction tasks. This is a significant limitation which this dissertation attempts to address. Modern Graphics Processing Units (GPUs) have become fully programmable and have spawned the field known as General Purpose GPU (GPGPU) processing. Using this technology it is possible to of- fload certain types of tasks from the Central Processing Unit (CPU) to the GPU. GPGPU processing is designed for problems that have data parallelism. This means that a particular task can be split into many smaller tasks that can run in parallel, the results of which and are not dependent upon the order in which the tasks are completed. Therefore to properly make use of both CPU parallelism and GPGPU processing a 3D reconstruction algorithm with data parallelism was required. The selected algorithm was the Patch-Based Multi-View Stereopsis (PMVS) method, proposed and implemented by Yasutaka Furukawa and Jean Ponce. This algorithm uses small oriented rectangular patches to model a surface and is broken into four major steps: Feature detection; feature matching, expansion and filtering. The reconstructed patches are independent and as such the algorithm is data parallel. Some segments of the PMVS algorithm were programmed for GPGPU and others for CPU parallelism. Results show that the feature detection stage runs 10 times faster on the GPU than the equivalent CPU implementation. The patch creation and expansion stages also benefited from GPU implementation. Which brought an improvement in the execution time of two times for large images, and equivalent execution times for small images, when compared to the CPU implementation. These results show that the use of GPGPU and CPU parallelism can indeed improve the performance of this 3D reconstruction algorithm.
Volumetric reconstruction of rigid objects from image sequences.
(2012) Ramchunder, Naren.; Naidoo, Bashan.
Live video communications over bandwidth constrained ad-hoc radio networks necessitates high compression rates. To this end, a model based video communication system that incorporates flexible and accurate 3D modelling and reconstruction is proposed in part. Model-based video coding (MBVC) is known to provide the highest compression rates, but usually compromises photorealism and object detail. High compression ratios are achieved at the encoder by extracting and transmit- ting only the parameters which describe changes to object orientation and motion within the scene. The decoder uses the received parameters to animate reconstructed objects within the synthesised scene. This is scene understanding rather than video compression. 3D reconstruction of objects and scenes present at the encoder is the focus of this research. 3D Reconstruction is accomplished by utilizing the Patch-based Multi-view Stereo (PMVS) frame- work of Yasutaka Furukawa and Jean Ponce. Surface geometry is initially represented as a sparse set of orientated rectangular patches obtained from matching feature correspondences in the input images. To increase reconstruction density these patches are iteratively expanded, and filtered using visibility constraints to remove outliers. Depending on the availability of segmentation in- formation, there are two methods for initialising a mesh model from the reconstructed patches. The first method initialises the mesh from the object's visual hull. The second technique initialises the mesh directly from the reconstructed patches. The resulting mesh is then refined by enforcing patch reconstruction consistency and regularization constraints for each vertex on the mesh. To improve robustness to outliers, two enhancements to the above framework are proposed. The first uses photometric consistency during feature matching to increase the probability of selecting the correct matching point first. The second approach estimates the orientation of the patch such that its photometric discrepancy score for each of its visible images is minimised prior to optimisation. The overall reconstruction algorithm is shown to be flexible and robust in that it can reconstruct 3D models for objects and scenes. It is able to automatically detect and discard outliers and may be initialised by simple visual hulls. The demonstrated ability to account for surface orientation of the patches during photometric consistency computations is a key performance criterion. Final results show that the algorithm is capable of accurately reconstructing objects containing fine surface details, deep concavities and regions without salient textures.

Browse

Browsing Masters Degrees (Computer Engineering) by Author "Naidoo, Bashan."