Pose Estimation Of Intelligent Vehicles Based On Monocular Visual Geometry

With the development of intelligent vehicle and visual sensor,vision-based pose estimation technology has gradually become one of the key technologies in the field of intelligent vehicle.At the same time,the researches on vision-based pose estimation are challenging due to large amounts of information collected by vision sensor and complicated practical scenes.This paper studies the monocular pose estimation method based on visual geometry with consideration of the characteristics of vehicular planar motion.According to the planar motion of intelligent vehicles,this paper simplifies the geometric model with the introduction of planar motion constraint,which is based on the existing two-view geometric model.Based on the simplified model,the relative pose between two views can be solved with the associated feature points in the two images.Furthermore,this paper designs a two-point algorithm to decouple the rotation from the relative pose and recover the rotation inde-pendent from translation.After that,the translation can be estimated with the rotation information.Because the two-point algorithm is a closed-form solver,the pose can be recovered stably and ef-ficiently with at least two pairs of associated feature points.The proposed algorithm can solve the pose with small parallax and even pure rotation,while epipolar geometry model is prone to degenerate.Synthetic data are designed to validate the proposed algorithm with the classical pose estimation algorithm.The results show that the proposed algorithm has the best robustness to noise of image feature points and the greatest adaptability to non-planar motion in most cases.As for the non-planar motion of vehicle in practice,the parameters of 6-DoF motion model are partly simplified with approximate treatment in this paper.And the optimization problem is constructed with the two-view geometric model which is solved by iterative optimization method.With the proposed two-point algorithm and the approximation method for the motion model,this paper improves the existing random sample consensus(RANSAC)algorithm to achieve accurate and robust pose estimation in practical scenes.Moreover,based on the classical visual odometry system framework,the visual odometry with the proposed algorithms is designed.Experiments with the simulation environment,benchmark and real vehicle platform are carried out to validate the performance of the visual odometry and compared with the method based on five-point algo-rithm.The experimental results indicate that the proposed method is superior to the method based on five-point algorithm with regard to accuracy and efficiency.