室内环境下基于拟大脑皮层模型的主动式视觉SLAM研究

To address the prolem that current active-mode SLAM systems have limitations on autonomy and intelligence, a novel active-mode visual SLAM framework is proposed in this project. The study of this new framework is based on the memeroy-prediction theory, which presents the basic principle of the neocortex, and its related simulated neocortex model (SNM). Firstly, after the appearance image sequceses of environments are constructed, map building process is designed and implemented by incorporating learning ability of SNM and appearance based mapping technology. The relationship between the data association problem and the inference capability of SNM is explored, and the data association algorithm is also proposed using Bayesian theory and machine learning method. Secondly, a behavior module is developed and inserted into the SNM for improvement. With this improved SNM, on the basis of the prediction ability of SNM the mechanism of sensory-motor coordination is investigated by applyng Baysian programming and dynamic field theory, and furthermore action selection and localiztion algorithms related to the sensory-motor coordination are developed. Finally, mapping, localization and action selection are integrated to build the whole active-mode visual SLAM, and without any priori information the method of action selection and performance evaluation strategy of this active-mode visual SLAM are further studied. The results of this project are promising and they will theoretically and technically provide a new idea and new way for the open research issue of active-mode SLAM.

针对现有主动SLAM的自主性与智能水平存在一些局限性的问题,本项目以提高主动SLAM的智能性为目标,根据描述大脑皮层工作原理的记忆?预测理论及相关的拟大脑皮层模型,提出一种新颖的主动式视觉SLAM系统。该系统的研究以拟大脑皮层模型为基础, 首先建立环境的外观图像序列,然后将模型的学习功能与外观制图方法相结合实现地图创建;探索拟大脑皮层模型的推理功能与数据关联问题的关系,根据贝叶斯理论与机器学习的方法提出相关算法。其次引入行为功能模块对拟大脑皮层模型加以改进,在此基础上利用贝叶斯规划方法与动态场理论,探究基于拟大脑皮层模型预测功能的感知-运动协调机理,设计与其相关的动作选择和定位算法。最后整合制图、定位与动作选择各环节构建完整SLAM系统,进一步研究无先验信息时的动作选择方法与SLAM系统性能评估策略。本项目以期拓展和丰富主动SLAM研究,在理论和技术上为解决该问题提供一种新思路与新方法。

针对现有解决视觉SLAM问题的方法存在一些局限性,本项目以提高机器人的自主能力为目标,提出了利用层级实时记忆(Hierarchical Temporal Memory, HTM)模型解决视觉SLAM问题的方法。该方法根据视觉SLAM的特点设计了具体的HTM模型的结构与参数；利用公众数据集和自建数据集提取图像的局部特征描述算子以训练生成外观制图技术中的视觉词汇表；根据词汇表将机器人运动过程中采集的图像用局部特征描述算子表示，并将描述算子映射为视觉单词频率矢量，多个视觉单词频率矢量构成具有时间序列特性的视觉单词频率矢量集，将该矢量集输入HTM 网络实现环境地图的学习与创建及环路场景的推断识别；采用 HTM模型的推理功能确定了数据关联定性方法，并结合基于偶图的定量数据关联方法实现一定程度的定量数据关联评价分析；将典型物体的图像序列输入HTM模型，经HTM模型推理并预测输出机器人遇到这些典型物体时所采取的典型行为策略，从而实现简单感知-运动整合；将机器人在运动中获取的图像输入视觉HTM网络模型得到相应的场景图像类，通过标量编码器将该图像类与深度和运动信息一起编码为稀疏分布表达形式的矢量序列，并将该序列作为输入训练感知-运动HTM网络模型，得到存储了图像、深度信息和运动数据之间转换关系的两个级联的HTM网络模型，从而实现高级感知-运动整合。高级感知-运动整合方法应用到结合了SLAM功能的示范式自主导航中，效果良好。本项目完成了研究内容，达到研究目标，为视觉SLAM 研究提供一种新思路与新方法。