多目鱼眼相机系统的视觉里程计解决方案

Visual odometer solution for a multi-fisheye-camera system

多目鱼眼相机系统具有可视范围广、成本低廉等诸多优点,成为自动驾驶系统中常用的传感器配置。但其成像畸变大、视角差异明显等特性,使得鱼眼相机间图像数据关联较为困难,大大限制了视觉里程计算法的精度。本文设计了一种可以用于自动驾驶场景下多目鱼眼相机系统的里程计解决方案及可以用于多目鱼眼相机系统的前后端算法,并实现两种鱼眼相机间数据关联模块:一是基于光度的数据关联方法,可以在光照差异明显或存在遮挡等条件下获得稳定的关联结果并纠正系统的尺度偏差;二是基于相机模型变换和描述子匹配的数据关联方法,可以加入优化系统提升里程计精度。实验表明,本文的解决方案可以达到平均约5‰的相对位置误差水平,可以满足自动驾驶场景对里程计的精度要求。

Multi-fisheye-camera is commonly used in autonomous driving system because of its wide visible range and low cost.However,the distortion and view angle difference make it difficult to establish good data association between fisheye cameras.This leads to the limitations on the accuracy of multi-fisheye-camera odometer.In this paper,we design a complete solution for visual odometer with multi-fisheye-camera configuration and presented a complete algorithm.Also,two innovative methods are proposed to establish the data association between fisheye cameras:one is a robust photometric method that enables scale correction even when illumination difference or occlusion exist;the other is a descriptor matching method based on warping model that can be integrated into the odometer system to boost accuracy.The experiments show that our solution can achieve high precision of about 5‰relative position error on average.The accuracy meets the requirement on an odometer solution for autonomous driving.