动态场景中的单目视觉里程计  

Monocular Vision Odometer in Dynamic Scene

作　　者：傅彬[1] 金棋 FU Bin;JIN Qi(Shaoxing Vocational&Technical College,Shaoxing 312000,China)

机构地区：[1]绍兴职业技术学院,浙江绍兴312000

出　　处：《电视技术》2022年第4期18-24,30,共8页Video Engineering

基　　金：浙江省产学合作协同育人项目“‘双师型’人工智能技术师资培育项目”(No.CX3202001)。

摘　　要：自动驾驶可以减少交通事故,提高出行效率,是当前研究的热门话题。自动驾驶最关心的问题之一是自身车辆定位、静态障碍物位置以及运动物体位置和朝向估计。对此,提出一个车载环境下的视觉里程计系统,此系统可以估计摄像机位姿,建立三维场景,利用卷积神经网络Mask R-CNN分割区分出背景和可能运动的物体。在惯性传感器单元(Inertial Measurement Unit,IMU)预测出摄像机运动的基础上,系统将可能运动的物体区分为正在运动的物体和当前静止的物体,根据静止的背景和当前静止物体上的特征点定位、三维重建及估计地面方程,提高视觉同步定位与建图(Simultaneous Localization and Mapping,SLAM)鲁棒性。在KITTI车载数据库上的实验表明,提出的方法可以精确地估计自身运动,重建静态三维场景。Autonomous driving can reduce traffic accidents and improve travel efficiency,which is a hot topic in current research.One of the most concerned issues in automatic driving is the positioning of the own vehicle,the location of static obstacles and the location and orientation of moving objects.This paper proposes a visual odometer system in a vehicle environment,it can estimate the pose of the camera and establish a three-dimensional scene.The system uses Mask R-CNN to distinguish the background and possible moving objects.Based on the prediction of the camera movement by the Inertial Measurement Unit(IMU),the possible moving objects are distinguished into moving objects and currently stationary objects.According to the stationary background and the feature point of currently stationary objects,this system can locate and three-dimensional reconstruct and estimate the ground equation,the robustness of visual Simultaneous Localization and Mapping(SLAM)is improved.Experiments on the KITTI vehicle database show that the method proposed in this paper can accurately estimate its own motion to reconstruct a static 3D scene.

关 键 词：卷积神经网络 图像分割 自动驾驶 目标检测 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]