基于分布式视觉的移动机器人大规模室内环境认知方法研究

The visual sensor has the advantages of rich information and low cost. Visual navigation has become a hot topic of robotic research. The effective visual field of vehicle visual navigation which is widely used is limited, and it is difficult to obtain and maintain a reliable model of dynamic environment of wide range. On the other hand, probabilistic navigation is with poor adaptability which relies on global metric model. If the robots want to realize the behavior of complex intelligence, it must have the ability to memorize and learn the environment, building a cognitive map of the environment to meet the robustness of environment modeling and localization. The project aims to explore and simulate human cognition mechanism for the intelligent service robot, proposes a distributed vision system architecture which is inspired by biological compound eye, make studies on the self-organized extraction process of the environmental features, learning method of spatio-temporal experience and multi-source heterogeneous information fusion algorithm, raises environmental perception to environmental cognition, makes an effective solution to key issues of traditional SLAM problem such as data association and the closed-loop detection. On this basis, an experimental system of distributed vision-based mobile robot for environment cognition is completed to verify the effectiveness and feasibility of the proposed method, which lay a good foundation to promote the development and application of intelligent service robots.

视觉传感器具有信息丰富、成本低廉等优点，视觉导航已经成为机器人研究的热点。目前广泛应用的车载视觉导航方式有效视野有限，较难获取和维护大范围动态环境的可靠模型；另一方面，流行的概率导航依赖于全局度量模型，适应性较差，机器人想要实现复杂智能行为，必须具备对环境的记忆和学习能力，构建环境的认知地图，以适应移动机器人自主导航的鲁棒性环境建模与定位要求。本项目旨在智能服务机器人方面探索和模拟人的认知机制，提出一种仿生物复眼的分布式视觉系统架构，对环境路标自组织提取、时空经验的自主学习方法以及多源异构信息的融合算法进行研究，将环境感知提高到环境认知，有效解决传统SLAM问题中数据关联、闭环检测等关键问题。在此基础上，完成一套基于分布式视觉的移动机器人环境认知实验系统，对所提方法的有效性和可行性进行实验验证，为促进智能服务机器人的发展和应用打下良好基础。

视觉传感器具有信息丰富、成本低廉等优点，视觉导航已经成为机器人研究的热点。目前广泛应用的车载视觉导航方式有效视野有限，较难获取和维护大范围动态环境的可靠模型；另一方面，流行的概率导航依赖于全局度量模型，适应性较差，机器人想要实现复杂智能行为，必须具备对环境的记忆和学习能力，构建环境的认知地图，以适应移动机器人自主导航的鲁棒性环境建模与定位要求。本项目提出了一种仿生物复眼的分布式视觉系统架构，对环境路标自组织提取、时空经验的自主学习方法以及多源异构信息的融合算法进行了研究，有效解决传统 SLAM 问题中数据关联、闭环检测等关键问题，使得机器人能够在人机并存的环境中规划出时空意义上的最优导航轨迹。