Visual Odometry with Reduced-Iterative Optimization Method for Automobile

Abstract

Visual odometry (vehicle motion estimation) is a fundamental technology for applications based on vision. Most applications require the latency to be very short. Methods/Statistical analysis: That is, it is necessary to reduce the computational load or the number of data points included in any calculation. However, this makes algorithms sensitive to noise, so most researchers have focused on methods to reduce their computational complexity instead.Findings: In this paper, we propose an improved visual odometry method for automotive applications. Generally, the optical flows of visual odometry algorithms have many outliers when applied to the automotive field; therefore, most approaches select N arbitrary points and then perform Random Sample Consensus. The error distance is calculated over the entire dataset, and inliers are selected based on a threshold. However, this increases the computational complexity, and the number of iterations required of this type of iterative execution over N points increases with N.Improvements/Applications: In this paper, we decrease this computational cost by improving execution time. Our method is a modification of the efficient perspective-N-point method, which uses four representative control points.