OUTLIERS DETECTION IN THE STEREO-VISUAL ODOMETRY BASED ON HIERARCHICAL CLUSTERIZATION

Abstract

This paper presents an approach to stereo-visual odometry without explicitly calculating optical flow. Visual odometry is a method of obtaining navigation information by processing a sequence of frames from onboard cameras. There are two approaches to processing video information - using well-localized areas of the image - feature points and using all high-contrast pixels - a direct method. The direct method works by using the intensities of all high-contrast pixels in the image, making it possible to reduce the computational complexity spent searching, describing, and matching feature points and increasing the accuracy of motion estimation. However, methods of this class have a drawback - the presence of moving objects in the frame significantly reduces the accuracy of the estimation of motion parameters. Outlier detection techniques are used to avoid this problem. Classical methods for detecting outliers in input data, such as RANSAC, are poorly applicable and have high computational costs due to the computationally complex function of rating hypotheses. This work aims to describe and demonstrate an outlier detection approachbased on the hierarchical clustering algorithm, which selects the statistically most probable solution, bypassing the stage of rating each hypothesis, which significantly reduces the computational complexity. For hierarchical clustering, a measure of the distance between hypotheses with low sensitivity to errors in estimating motion parameters is proposed. An extension of the stereo-visual odometry algorithm is also proposed to work in more complex visibility conditions due to the transition from the intensity representation of the image to a multichannel binary one. Transforming an image to a multichannel binary representation gives invariance to changes in image brightness. However, it requires modification of nonlinear optimization algorithms to work with binary descriptors. As a result of the work, it has been shown that the proposed outlier detection algorithm can operate in real-time on mobile devices and can serve as a less resource-intensive replacement for the RANSAC algorithm in problems of visual odometry and optical flow eviction. Qualitative metrics of the proposed solution are demonstrated on the KITTI dataset. The dependences of the performance of the algorithm on the parameters of the algorithm are given.