双目红外显著目标提取及其硬件实现

发布时间：2018-05-14 07:31

本文选题：双目立体夜视系统 + 极线校正　；参考：《南京理工大学》2017年硕士论文

【摘要】：随着夜视技术的发展,红外探测器被广泛地应用在无人驾驶、机器导航、虚拟现实及医学图像处理等领域。由于单红外相机只能采集图像的表层信息,不利于捕捉深度信息,因此利用双目立体夜视技术来实现图像场景的三维重建是发展的趋势。本文基于双目立体夜视原理,研究了极线校正、显著目标检测,设计并搭建双目红外显著目标提取硬件平台。双目立体夜视利用左右红外相机进行图像采集,由于存在装配误差,不能确保两路相机处于完全水平状态,导致图像对不满足极限约束等缺陷,本文研究了基于单一变换矩阵的极线校正方法。该方法利用左图作为基准图,对右图像进行线性变换,使得变换后的右图像与原始左图像为校正后的图像对。针对双目红外显著目标检测,本文提出并研究了基于LARK的局部特征的全局对比度显著性检测方法。该方法利用局部特征协方差矩阵作为特征值进行显著性检测,通过引入空间加权信息和亮度权值信息,将局部显著性检测扩展到全局显著性检测。该方法不仅运用了图像的特征信息来衡量显著性,同时考虑了全局对比度和空间相干性来区分显著目标和背景。针对基于背景信息的显著性检测,本文提出并研究了一种有效的距离估算模型。该模型利用像素之间的距离来提炼边界超像素,同时利用边界对比度,精化边缘信息,最终达到改善显著模型检测效果。针对系统硬件化实现,设计并搭建了基于多核DSP6678和FPGA为核心的硬件平台。将极线校正及显著提取等算法集成到硬件系统中,并在不同场景下进行实验,实验结果表明,该系统能够达到实时性处理,并且能很好的提取出显著目标信息。
[Abstract]:With the development of night vision technology, infrared detectors are widely used in the fields of pilotless, machine navigation, virtual reality and medical image processing. Because the single infrared camera can only collect the surface information of the image, it is not conducive to capture depth information. Therefore, the use of binocular stereo night vision technology to realize the three-dimensional reconstruction of the image scene is the development. Based on the principle of binocular stereo night vision, this paper studies the pole line correction, the significant target detection, and the design and construction of the double eye infrared significant target extraction hardware platform. The binocular stereo night vision uses the left and right infrared cameras for image acquisition. Because of the existence of assembly errors, the two cameras can not ensure that the two cameras are in a complete level state, resulting in the image correct. In this paper, the pole line correction method based on the single transformation matrix is studied in this paper. This method uses the left image as the reference map, and makes the linear transformation of the right image, making the right image after the transformation and the original left image as the corrected image. The paper proposes and studies the LARK based on the binocular infrared explicit target detection. The local feature covariance matrix is used as the eigenvalue to detect the feature value. The local saliency detection is extended to the global saliency detection by introducing the spatial weighted information and the weight information of the luminance. The method not only uses the feature information of the image to measure the obvious. In order to distinguish significant targets and backgrounds from global contrast and spatial coherence, an effective distance estimation model is proposed and studied in this paper for the significance detection based on background information. This model uses the distance between pixels to extract boundary super pixels, and uses boundary contrast to refinement edge information. According to the hardware realization of the system, a hardware platform based on multi core DSP6678 and FPGA is designed and built. The algorithm of pole line correction and significant extraction is integrated into the hardware system, and the experiments are carried out in different scenes. The experimental results show that the system can achieve real-time processing, and the results show that the system can be processed in real time, and It can be very good to extract the significant target information.

【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN21;TP391.41

【相似文献】