一种改进的BEMD阈值去噪算法在单目视觉定位中的应用

Application of improved BEMD threshold denoising algorithm in monocular visual odometer

针对单目视觉定位系统获取的图像含有较大噪声的问题,提出了一种改进的二维经验模态分解(BEMD)阈值去噪算法。首先,通过构建噪声压缩图像消除BEMD分解的位置敏感性误差;然后,通过BEMD将构建的噪声压缩图像分解成一系列本征模态函数,并基于?2准则和概率密度函数区分出噪声主导本征模态函数和信号主导本征模态函数,使用软阈值技术实现噪声主导IMF的去噪。最后,将所提算法应用于单目视觉定位中,并与小波去噪和BEMD-DT去噪算法进行对比。对比结果表明,在转弯和直线运动两种情况下的东向位置误差、北向位置误差和航向角误差方面,该算法较Sym4小波去噪算法分别平均改善了74%、64%、54%,较Db6小波算法分别平均改善了60%、48%、39%,较BEMD-DT算法分别平均改善了73%、96%、84%,显示了所提算法的优势。

Aiming at the problem that the image acquired by monocular visual positioning system contains large noises, an improved threshold denoising algorithm for bidimensional empirical mode decomposition(BEMD) is proposed. First, the position sensitivity error of the BEMD is eliminated by constructing the noisecompressed image. Then, the noise-compressed image is decomposed into a series of intrinsic mode functions(IMFs) by the BEMD, which are separated into signal-dominant IMFs and noise-dominant IMFs by using a similarity measure based on ?2-norm and a probability density function, and a soft thresholding technique is used adaptively to remove the noise inherent in noise-dominant IMFs. Finally, the algorithm is applied to the monocular visual location and compared with the wavelet denoising algorithm and the BEMD-DT one. The comparison results show that, in both cases of turning and linear motion, the east, north and heading errors of the proposed algorithm are improved by 74%, 64% and 54% respectively compared to that of the Sym4 wavelet denoising algorithm, and are improved by 60%, 48% and 39% respectively compared to that of the Db6 wavelet denoising algorithm, and are improved by 73%, 96% and 84% respectively compared to that of the BEMD-DT algorithm, which show the superiority of the proposed algorithm.