On the Comparison of Mono Visual Odometry Front End in Low Texture Environment

Visual odometry is the process of determining the position and orientation of a vehicle using associated camera images. As we known, the quality of outcomes of the six degrees-of-freedom (DoF) poses created by visual odometry play a decisive role in autonomous location, map creating, and path planning in SLAM system. While different approaches for handling the monocular visual odometry have been used in practice, but few previous studies have been carried out to systematically analyze their differences, especially in a repeat scene or a low texture environment which detect a small amount of feature points. In this paper, we present the comparative analysis of ORB feature detection and matching, and Shi-Tomasi detection and optical flow matching in visual odometry front end process. We briefly introduce the commonly used Perspective-n-Point (PnP) methods and experimentally compare three PnP approaches: based on linear method DLT, EPnP, and Bundle Adjustment (BA) which based on nonlinear optimization method. We built a simulation data set to evaluate those PnP approaches, and finally sum up an optimal combination of visual front-end in the area of low texture environment based on the calculation efficiency, reliability and accuracy.