Enhanced Monocular Visual Odometry: A Robust Appearance-Based Method for Accurate Vehicle Pose Estimation

Abstract

Monocular Visual Odometry (MVO) is a fundamental element in autonomous navigation systems, providing vehicles/robots with the capability to estimate their positions by analyzing visual images from a single camera. This work delves into a pure appearance-based MVO algorithm that estimates the vehicle displacement and orientation between consecutive image frames alone, without using an Inertial Measurement Unit (IMU) sensor. The proposed method comprises four stages: ground spatial calibration, vehicle displacement, orientation estimation modules, and an actual vehicle heading estimation module. In the first stage, the image pixel coordinates are converted into world coordinates through ground spatial calibration. In the second stage, cross-correlation-based template matching is performed between two successive image frames and vehicle displacement is computed using the obtained world coordinates. Next, the orientation of the matched template is estimated along the ‘u’ and ‘v’ axis of the image. Subsequently, the actual vehicle heading is computed in the fourth stage with respect to the global coordinate system to estimate the vehicle pose. Experimental evaluations demonstrate the superior performance of the developed MVO algorithm compared to existing appearance-based methods that additionally utilize IMU to obtain orientation. When the vehicle is driven for a distance of 1406.35 meters, the average percentage distance error obtained is 1.41%, thereby highlighting the improved performance of the MVO algorithm in terms of higher accuracy and efficacy in real-world applications.

Keywords

• Autonomous navigation
• ground spatial calibration
• template matching
• visual odometry