Optoelectronic Target Detection Methods Based On GPUs

With the rapid development of both hardware and software for computer graphics, the reality of simulated images has been improved a lot in the last decade. Introducing the graphics processing unit (GPU) technique to the optoelectronic target detection (OETD) system will benefit in two aspects, on one hand the rendering power of GPUs can en-hance the target simulation, on the other hand the parallel computing ability of GPUs can accelerate the speed of the detection methods. Therefore, a complete chain of target de-tection can be constructed based on GPU as "Simulation-Detection-Feedback", which can be used to test or evaluate the performance of OETDS. The GPU-based OETDS has the following features:the low cost for experiments, the short time for development, no limitation by the environment and the easy operation of the system. Meanwhile, as the optoelectronic device has more powerful ability to capture target information, the amount of collected data becomes larger and larger. As a result, how to efficiently conduct the target detection plays an important role in the whole OETD system. It determines the performance and feasibility of the OETD system. Nonetheless, GPU’s massively parallel high performance computing (HPC) ability is a promising solution to improve the perfor-mance of OETD system. Thus, the GPU-based OETD system is of great interest in the world.Focusing on the powerful rendering and parallel computing ability of GPUs, this thesis describes a feasible method to improve the performance and enhance the reality of optoelectronic target simulation, and presents a parallel design for high performance implementation of the OETD method. The gist of the thesis is as follows:The first section introduces the main structure of GPUs’ hardware and the parallel computing mechanism. CUDA programming model is briefly described with some opti-mal techniques for parallel design with GPUs. By analyzing the render pipeline in a GPU, the key technology and the performance bottlenecks are pointed out. In addition, the ba-sic physical principles for illumination models are presented in this section. Both field and simulated experiments have been conducted to verify the practicality and efficiency of the proposed method. Finally, recommendations are made to select proper GPUs for different applications.The second section describes three methods for different optoelectronic target simu-lation. Firstly, taking advantage of PRISSE (Physical Reasonable Infrared Scene Simula-tion Engine), a wavelength based BRDF model has been proposed to depict the reflection property of a mid-wave infrared target. Secondly, a digital optoelectronic imaging ren-der (DOEIR) platform is constructed to simulate the complex radiative transfer between multiple targets. In DOEIR, GPU is adopted to implement the ray tracing algorithm. Fi-nally, for hyper-spectral target simulation, a linear spectral mixture model is implemented based on GPU and Dirichlet distribution is introduced to simulate the abundance for each endmember.The third section shows the GPU-based implementation of infrared images back-ground prediction algorithm. Considering the time consuming convolution in the back-ground prediction algorithm, separable convolution template is proposed to decompose the filter matrix into vectors. Moreover, the convolution for separated template is parentally processed in GPU, which gains a significant speedup. Embed GPU is also used to test the performance of the proposed separable convolution background predic-tion method.The forth section depicts the efficient hyperspectral target detection method pro-posed in this thesis. The hyperspectral image usually involves with huge matrix opera-tions (transpose, inverse, decomposition and et al.), therefore, it is difficult for target detection algorithm to process hyperspectral images in real time. By analyzing the flow and bottlenecks of hyperspectral image processing, a GPU accelerated implementation of identifying subspace of hyperspectral image is presented. Besides, matrix inverse up-dating and downdating technique is adopted to accelerate the unsupervised non-negative constrained least squares method to detect hyperspectral target. The performance has been evaluated by real and simulated hyperspectral images, experiments show that the proposed method can process the hyperspectral data collected by AVIRIS in real time.