Line Segment Probability Alignment Direct Visual Odometer

Abstract

The direct method is to compare the pixel difference between two frames of images. However, the gray image and photometric error only keep convexity in a small range. Therefore, when there is a large displacement of the camera position, the system may fall into a suboptimal local minimum. Semantic features can solve this problem, but they can only be run in scenarios where semantic features are known. Line features can extract edge information and have good convexity for large camera displacement. In addition, it is easy to adapt to different scenes. In this letter, we propose a line feature probability matching visual odometer. We have built a lighter line feature probability estimation network, which can be deployed on platforms the limited computational power. A joint error function based on gray image and line feature probability is constructed, which has better reliability than photometric error. We experiment the proposed method on public indoor and outdoor datasets, and the results show that the joint feature probability error function is significantly improved than the original method.