Performance Enhancement Step for Motion Estimation via Feature-based Image Matching

Most of the complicated and sophisticated tasks in visual robotics applications usually build upon the image matching step as matching images of the same scene can provide important information (e.g., camera motion). Image matching is generally done via extracting and matching some distinctive points via their feature vectors. This procedure generates some mismatched points due to imperfections. Mismatched points are called outliers and identified via probabilistic methods. Since the probabilistic methods work iteratively, they generally occupy a large portion of the computational cost of the whole image matching pipeline. In this paper, we present a simple yet efficient algorithm that is employed for eliminating the outliers aiming at reducing the total number of iterations needed in the probabilistic methods. Our method is motivated by the common way of visualizing the established matches among images. We tile images together and search for parallel lines connecting correspondences. We present extensive computational and comparative experiments using both simulated data involving along with real images and using a real dataset.

1. Introduction

Thanks to the advancements in sensing technology, camera-carrying mobile platforms have become more and more accessible and available for a vast variety of engineering and science disciplines. This also leads to scientific developments in mainly computer vision, image processing, and machine learning areas. Without a doubt, image matching (or registration) has secured its position and necessity at the core of many more sophisticated tasks. It is defined as a procedure of overlaying two images that have some overlapping area. In order to overlay images, the coordinate transformation (motion) between their coordinate frames (top-left corner as origin) needs to be calculated. The success of several different high-level methods in computer vision and robotics (e.g., mapping, 3D reconstruction, localization, and similar others) relies on image matching. Image registration methods are mainly categorized in three categories [22]; Optical Flow [11, 19], Fourier Transform based [16], and Feature-based. Over last two decades, developments on feature point detection and description (usually Scale invariant feature transform (SFT) [12] and Speeded up robust features (SURF) [2]) made it possible to compute the transformation between images even under extreme cases (e.g., various scale and viewpoints changes). These advancements direct researchers to use Feature-based methods more. Different deep-learning-based methods have been also proposed for feature detection, and matching (e.g., [13, 21]) and comparative benchmarking was presented in [1]. The D2-Net framework for joint detection and description of local features was proposed in [6]. Its performance in localization tasks outperforms the other methods while in image matching, it has some limitations.