时变环境下基于人类记忆机制的长期SLAM技术研究

At present, most of the Simultaneous Localization and Mapping (SLAM) researches are concerned only with the filtering method, dynamic objects processing and et al, and are limited to realize short-term running. This project aims at studying on long-term SLAM technologies in time varying environments by introducing human memory model, in order to provide new ideas about the design and application of long-term SLAM system. This project will mainly focus on the mapping, localization and learning methods of long-term SLAM, and include the following aspects. Firstly, map building method in long-term SLAM is researched with the human memory mechanism for time varying environments. Minimum feature set is used to characterize the same scene in long term memory by evaluating the stability of the features, and the eventual spatial-temporal map based on the scene appearance samples is eventually established. This method can not only reflect changes in the appearance of the scene over time, and alleviate the huge storage capacity needed by the SLAM system long-running. Secondly, a long-term SLAM algorithm integrating the memory model is designed. The optimal pose graph is calculated by manifold learning optimization with constraint relations by the scene registration and closed-loop detection. Finally, the learning of the long-term SLAM system is designed to deal with learning and recall of scene samples, and to be adaptable for time varying environments.

针对目前大部分同时定位与建图（SLAM）研究仅关注滤波方法、动态目标处理等方面且局限于系统的短期运行，本项目引入人类记忆模型研究时变环境下的长期SLAM技术，为长期SLAM系统的设计和应用提供新思路。本项目将围绕长期SLAM中的地图构建、SLAM算法框架和学习机制三个方面展开，主要包括：1）针对时变环境，结合人类记忆机制研究一种面向长期SLAM的地图构建方法。通过评估特征的稳定性获取能以最小特征集表征同一场景的长期记忆，最终建立基于场景外观样例的时空地图，能反映场景外观随时间发生的变化，并缓解系统长期运行出现的存储量问题；2）研究融合记忆模型的长期SLAM算法，通过场景配准和闭环检测获取位姿图的约束关系，采用流行优化方法求解最优位姿序列，同时更新地图；3）设计长期SLAM系统的学习机制，解决时变环境下对场景样例的学习和回想，使系统对时变环境具有适应性。

本项目旨在研究时变环境下长期SLAM技术。主要内容包括：（1）针对群组图像特征的检测，研究了一种基于多示例学习算法的单幅和群组图像协同显著性检测方法；（2）研究了基于稀疏表达和尺度学习的目标跟踪方法，用于解决长期SLAM中动态目标的影响；（3）针对室内环境，研究了一种基于平面特征的鲁棒RGB-D SLAM方法；（4）针对异步传感器信息融合的问题，研究了一种基于平面共面约束的摄像头和激光融合的SLAM方法；（5）研究了一种基于对比散度-受限玻尔兹曼机深度学习方法，可用于长期SLAM系统中不同场景的分类。