基于认知地图的类脑行人室内导航方法研究

The state-of-the-art indoor navigation methods for pedestrian are applied in a limited area with specific infrastructure, signal environments, or devices. In order to realize a robust and non-constrain indoor navigation and localization in large-scale space with low cost sensors, inspired by the navigation mechanism of mammal brain, we propose a brain-like indoor navigation approach based on cognitive map. This approach targets on wearable body sensor network and main research contents involved in this project include the following: .1) In order to obtain the self-motion information from a wearable body sensor network , we propose a HWB PDR（Head-wrist-body Pedestrian Dead Reckoning）method, which estimates the heading and travel distance from multiple sensors attached on a pedestrian in a natural way. Using this method, we will achieve a non-constraint, non-linear, high degree of freedom, and robust self-motion estimation with low cost sensors. .2) In order to realize the cognition of an environment and memorize visited places, we propose a Hippocampus-like Spatial Cognition Model inspired by Hippocampus part in brain. Spatial feature extraction and coding mechanism is applied for spatial cognition to generate Cognitive snapshots. Inspired by the mechanism of “Place cell” generation, we design a model to produce a long-term memory for the visited places with unique features..3) We propose an Entorhinal Cortex-like Navigation Map Model, which imitates “Grid cell” in Entorhinal Cortex to organize spatial information and build “Cognitive Map”. An egocentric grid map and corresponding spatial transform theorems are designed. Self-motion information and Cognitive snapshots are projected onto the “Grid cells” of the grid map. Continuously projected Cognitive snapshots during a pedestrian walking are merged into a Cognitive map on top of the grid map. .4) For the complex path integration, we propose factor graph based solution, where prior information in the “Grid cells”, sparse and asynchronous environmental signals of the Cognitive snapshots, will be fused with self-motion information. The complex and non-linear navigation process is divided into a sequential of variable nodes and factor functions, which enable a plug-and-play solution with lower complexity. Finally, an optimal trajectory on the grid map of the pedestrian is estimated.

目前行人室内导航方法均局限在特定的小块区域，依赖特定的基础设施、信号环境或设备。为实现大范围、无约束、低成本、鲁棒的室内定位导航，受哺乳动物导航机制启发，本项目提出基于认知地图的类脑行人室内导航方法。该方法面向穿戴式体感网，1）提出头部、躯干、肢体多部位约束的行人运动模型，用于提高基于体感网的行人航向与里程估计精度；2）提出类海马体空间认知模型，模拟海马体进行特征抽取与特征编码，实现对空间信息的认知，模拟海马体“位置细胞”生成机制，形成对某个特定区域空间特征的长期记忆；3）提出类内嗅皮层导航地图模型，模仿大脑内嗅皮层中“网格细胞”建立网格地图框架，进行空间特征的组织与表达，构建认知地图；4）提出基于因子图的路径整合方法，融合先验的“位置细胞”信息、行人自身的运动信息、认知地图中的信号特征观测量，将在网格地图中实现路径的最优估计。

目前行人室内导航方法均局限在特定的小块区域，依赖特定的基础设施、信号环境或设备。为实现大范围、无约束、低成本、鲁棒的室内定位导航，受哺乳动物导航机制启发，本项目提出基于认知地图的类脑行人室内导航方法，分别从行人运动建模、空间认知模型构建、导航地图模型构建、路径整合四个方面展开研究。项目主要成果总结如下：.（1）针对行人运动建模，项目创新性地建立了混合现实运动识别方法，实现了在有限训练样本条件下的鲁棒运动识别模型和无约束场景下的连续运动分割框架，在Hospital数据集上Fm指标高于先进对比算法16.08%，为仿脑行人导航的深入研究奠定了基础。.（2）围绕着空间认知模型建模，重点研究了脑电信号认知、空间认知编解码、类脑导航机制深度神经网络模拟，并针对有效图像空间特征的提取的问题，提出了一种自适应特征提取优化策略，该策略可应用于大多数基于直接法视觉定位系统，从而提升系统的效率、准确性和鲁棒性。.（3）基于类脑导航的生物学基础，从内嗅皮层上的网格细胞和海马体上的位置细胞对常见导航任务进行建模分析，经过对感知快照的特征编解码，针对不同细胞模型、不同感知类型、不同尺度场景和不同导航任务进行了实验分析，实现导航认知地图模型的构建，为类脑导航实现提供环境感知基础。.（4）基于环境感知信号的稀疏性、不同信号的异步性、不同传感器的采样频率差异性、传感器组合的灵活性，项目融合多个模态信息，结合因子图/滤波算法，提出了DVT-SLAM、P3-LOAM、P3-VINS、P3-LINS等路径整合及定位算法，在一分钟以内将定位和姿态初始化误差分别降低到1米/1°以内。