【作者】 张文；

【导师】 刘勇；

【作者基本信息】 中国科学技术大学 ， 光学， 2017， 博士

【摘要】 自主导航是室内机器人实现自决行为规划的核心问题,相对复杂条件下的环境感知,基于多传感器融合成为新的发展趋势。本课题以机器人平台为工程应用背景,期望将深度学习引入人工智能领域,开展相关机器人技术在自主导航中的方法研究,开发集机械平台、嵌入式硬件、软件系统、SLAM算法、场景识别方法于一体的机器人综合系统框架,结合多传感器融合的环境信息,实时指导路径规划,不仅提供室内机器人自主导航的可行方法,对促进该领域语义地图的发展也具有一定积极意义。本文主要内容如下:研究并设计满足自主导航条件的机器人综合系统框架,包括集主动感知、语音输入、运动控制等的机械平台,集IO驱动、数学运算库、机器人控制、运动学和动力学模型等的嵌入式硬件结构,集环境感知与执行、移动终端、中间层连接、分布式架构、任务算法等的机器人系统。研究适用于ROS环境的多传感器融合SLAM算法:融合超声、RGBD、激光雷达数据,基于Gmpping建立二维栅格地图,通过自适应蒙特卡洛定位,导航时采用A*算法实施全局探索,并辅以动态窗口法进行局部避障。实验证明所创建地图分辨率为2cm,并能指导机器人进行室内导航。提出一种混合IMU的里程计实时校正算法:基于扩展卡尔曼滤波和互补滤波融合获取姿态角,修正机器人方位信息;定时比较里程计位移与二次积分加速计数据,防止机器人运动漂移及悬空。相关传感器经过零漂校准后,融合实验数据精度高、收敛速度快,并能校正里程计丢失现象。提出一种用于路径规划的方向A*算法:首先采用“视野线”平滑原则优化路径,消除锯齿效应并避免部分碰撞;其次应用“圆弧-直线-圆弧”转弯策略,避免机器人本体宽度影响;最后基于二叉堆加速算法,提升算法计算效率。仿真实验结果表明,方向A*算法满足平滑要求且能有效避免碰撞,其加速方法可平均提速4-7倍。同时,机器人在真实实验环境下实现安全自主导航,跟踪误差小于0.15m,验证了方法的可行性。提出一种基于语义建图的场景分类方法:首先映射深度信息构建二维栅格地图,自主规划场景识别路径;其次基于卷积网络建立场景分类模型,实时识别脱离特定训练;最后利用贝叶斯框架融合先验知识,修正错误分类并完成语义建图。实验结果表明,机器人能够进行全局自主探索,实时判断场景类别,并创建满足要求的三维语义地图。同时,在实际路径规划中,机器人可以根据语义信息改善导航行为,验证了方法的可行性。更多还原

【Abstract】 Autonomous navigation is the core issue of indoor robot’s self-determination behavior planning.Because there are so many problems in complex environment,the method which based on multi-sensor fusion has become a new trend.On the background of engineering application,and with the hope of introducing the deep learning into the field of artificial intelligence,we carried out some research on the using of related robot technologies in autonomous navigation.We developed an integrated robot system framework,including mechanical platform,embedded hardware,software,SLAM algorithms and so on.Combining with multi-sensor fusion environment information,we not only provided a solution to the common problems in the navigation tasks,but also promoted the development of semantic mapping.The main contents of this paper as follows:First,we developed an integrated framework of robot system,including a mechanical platform,an embedded hardware architecture and a robot software system.The mechanical platform consists of active perception,voice input,motion control and other modules.The embedded hardware architecture includes some hardware elements,such as 10 drive,mathematical operation library,kinematics library and dynamic model.The robot software system refers to the bootbot system,provided mobile terminal,intermediate connection,mission algorithms and so on.This integrated framework we designed can meet the conditions of autonomous navigation.Second,we investigated a SLAM algorithm based on multi-sensors for ROS environment.Firstly,we created the grid map based on Gmapping method,by mixing the data from ultrasound,RGBD camera and laser.Then,we used both the adaptive monte carlo location algorithm and A*algorithm for global exploration.Furthermore,the dynamic window method was applied to avoid local obstacles.The experiment results showed that the map had a resolution of 2 cm and can guide the robot for indoor navigation.Third,we proposed a real-time correction algorithm which fused the information of IMU.We used the Extend Kalman filtering and complementary filtering to acquire the position of robot,and by constantly comparing odometer displacement and quadratic integral accelerate data to prevent the movement and suspension of robot.The results indicated that the fusion data was stable and the accuracy was high for calibrating the odometer loss.Fourth,we proposed a kind of directional A*algorithm.Firstly,the "line of sight" solutions was used to smooth the path for getting rid of the zigzag effect and collisions.Secondly,the "arc-line-arc" turning methods was applied to avoid the width of the robot in path-finding.At last,some basic optimizations based on the binary heap were carried out to speed up the directional A*algorithm.Simulation and comparison results between the improved A*algorithm and traditional one showed that the method we proposed was more efficient.At the same time,the accelerating algorithm based on the binary heap made the path-finding 4-7 times faster.Moreover,a path planning and tracking test was carried out in lab environment,the results verified that the tracking precision can keep in a small range and the robot can run without collision when the navigation path was given by the proposed algorithm.Finally,we proposed the semantic mapping algorithm on a robot without environment-specific training.Firstly,the two-dimensional grid map was constructed to plan the recognition path of scene classes with depth information.Secondly,a state-of-the-art convolutional network was applied to recognition semantic classes without environment-specific training in real-time.At last,a Bayesian estimation framework,which incorporating prior domain knowledge,were carried out to smooth out spurious results in semantic map.Simulation and experiment results,which was used to evaluate the classification system on a robot in different places of our center,showed that the method we proposed was efficient.At the same time,the probability model based on Bayesian framework can fix error classification.Moreover,a path planning and navigation test was carried out in lab environment.The results verified that the robot can modulate its behavior with semantic information.更多还原