基于ConvNet路标语义时空关联与在线聚类的移动机器人视觉地点识别

Autonomous mobile robots, represented by unmanned vehicles, have significant civil and military value, while visual location recognition is one of the key technologies to realize visual-based autonomous navigation and localization. Nowadays, visual place description, remembering and matching technology is the main obstacle to the practical application of mobile robot visual location recognition under challenging environmental conditions. Comparing to methods with handcraft features or ConvNet-based global features, visual place recognition with ConvNet landmark features is able to achieve environment- and viewpoint- invariance. Despite that, it is still difficult to satisfy the practical application requirements in terms of the repeatability of detected landmarks and online scalability of place remembering. Therefore, we first establish a semantic spatio-temporal association model of landmark sequences to select ConvNet landmarks with high repeatability, and then propose a novel visual place description algorithm with stronger environmental condition invariance. Second, we analyze the online clustering characteristics of ConvNet landmark features in order to propose a visual place remembering and matching algorithm with better online scalability. Finally, we develop a visual place recognition method and system with both of environmental condition invariance and online scalability. The outcome of the project would make a contribution for the practical application of mobile robot visual location recognition, and further enrich and develop the research of visual place description, remembering and matching under complex conditions.

以无人车为代表的自主移动机器人具有重要的民用和军事价值，而视觉地点识别是其实现自主视觉导航定位的核心基础技术。当前，复杂环境条件变化下的地点描述与记忆匹配技术是制约移动机器人视觉地点识别实用化的主要障碍。与基于手工特征或ConvNet全局特征的方法相比，基于ConvNet路标特征的视觉地点识别能更好地兼顾环境不变性和视点不变性，但在路标检测的可重复性、地点记忆的在线可扩展性等方面仍难以满足实际应用要求。本项目拟首先建立序列路标语义时空关联模型，筛选可重复性高的ConvNet路标，研究环境条件不变性更强的地点描述算法；然后分析ConvNet路标特征的在线聚类特性，研究在线可扩展性更好的地点记忆与匹配算法；最后构建兼备环境条件不变性和在线可扩展性的视觉地点识别方法与系统。研究成果将为移动机器人视觉地点识别实用化创造条件，同时也将进一步丰富和发展复杂环境条件下视觉地点描述与记忆匹配的研究。

本项目针对复杂环境条件变化下移动机器人视觉地点识别实用化的难点，建立了序列图像ConvNet路标语义时空关联模型，提出了一种基于时空语义关联路标的视觉地点描述算法，筛选了可重复性高的路标，提高了视觉地点描述算法的环境条件不变性；利用机器人导航定位应用场景特殊性及ConvNet路标特征的语义时空相关性，提出了一种基于图卷积神经网络两阶段融合的视觉地点记忆与匹配算法，较好地兼顾了匹配效率和准确率；最终从整体上提出了兼备环境条件不变性的视觉地点识别方法，构建相应的软硬件原型系统，并进行了实测应用性能评估，验证了本项目所提方法可以满足复杂环境条件变化下移动机器人视觉地点识别实用化需求。本项目研究达到了为移动机器人自主视觉导航定位投入更加广阔的实际应用提供前沿基础理论与先进技术支撑的总目标。