基于光学成像模拟的视觉里程计方法

Visual Odometry Method Based on Optical Imaging Simulation

摘要:

针对视觉同步定位与建图(SLAM)算法前端漂移误差累积问题,提出基于光学成像模拟的视觉里程计方法。该方法利用稠密三维点云数据进行目标表面泊松重建及材质物理特性关联,依据光学物理特性和光线追踪原理构建基于物理模型的成像渲染引擎(PBRT),生成不同观测条件下的目标特性仿真图像；将目标特性仿真图像与光学相机拍摄图像进行配准与运动偏差恢复,并设计扩展卡尔曼滤波器(EKF)输出位姿状态最优估计值。通过原型系统研制与实验评估表明:该方法有效克服了传统方法漂移误差累积的问题,相较传统ORB-SLAM2算法前端定位精度提升了56%,为视觉里程计的设计提供了一种新的技术思路。

Abstract:

Aiming at the problem of front-end drift error accumulation of the V-SLAM algorithm, a visual odometry method based on optical imaging simulation was proposed. This method used the dense 3D point cloud data to perform Poisson reconstruction of the target surface and correlation of the physical properties of materials, and then built a physically based rendering engine (PBRT) according to optical physical properties and ray-tracing principle to generate simulated images of the target characteristics under different observation conditions. The simulated images of the target characteristics and the image captured by the optical camera were registered and the motion deviation was recovered, and the extended Kalman filter (EKF) was designed to output the optimal estimation value of the pose state. The prototype system development and experimental evaluation show that this method effectively overcomes the problem of drift error accumulation in the traditional methods and improves the front-end positioning accuracy by 56% compared to the traditional ORB-SLAM2 algorithm, which provides a new technical idea for the design of visual odometry.