基于深度滤波器优化的SLAM单目稠密重建

SLAM monocular dense reconstruction based on depth filter optimization

稠密重建问题是视觉同时定位与地图构建(SLAM)的重要环节,每一个像素点深度距离的准确测量对稠密重建都起到重要作用。在工业应用中,往往使用RGB-D相机进行稠密重建,但是RGB-D相机有一些量程、应用范围和光照的限制。因此,采用滤波器方式深度估计的单目相机,不仅可以保证SLAM实时性要求,同时还适用于室外、大场景等场合。针对高斯滤波算法存在稠密重建准确率不高的问题,提出了一种基于簇的均匀—高斯深度滤波算法,采用改进的滤波算法处理错误匹配的像素点,在正确处理外点数据的基础上,解决深度值错误估计、相邻像素深度值相差过大的问题。实验结果表明:改进型深度估计算法重建的稠密地图更加细致,且重建准确率提高了约30%。

Dense reconstruction is a significant component of visual simultaneous localization and mapping(SLAM).Accurate measurement of depth distance of each pixel point plays a crucial role in dense reconstruction.In industrial applications,the RGB-D camera is generally utilized for dense reconstruction;however,the RGB-D camera still remain certain unavoidable shortcomings,i.e.,small measuring range,narrow application scope,and high sensitivity to illumination.Hence,the depth estimation of the monocular camera with filter mode not only guarantees the real-time requirement of SLAM,but also be suitable for outdoor and large scenarios.A cluster-based uniform Gauss distributed depth filtering algorithm is used to overcome the difficulty that the dense reconstruction of the deep filtering under the Gauss distribution hypothesis with an unsatisfactory accuracy.The improved algorithm is adopted to deal with the mismatched pixels point,and also significantly robust against outliers.In addition,on the basis of correctly processing the outliers data,the problem of depth value estimation error and big difference of adjacent pixel depth value is solved.The experimental results show that the improved depth estimation algorithm is more delicate in reconstruction of dense maps,and the reconstruction accuracy is improved by about30%.