基于仿生双目视觉技术的高动态室外场景的即时定位与建图系统

The project is designed to make an intensive study of robust simultaneous localization and mapping (SLAM) system for high-dynamic outdoor applications based on bionic stereo vision. There is a succession for each part of the research, that is, firstly we set up the oculomotor control model for bionic eyes and thus take advantage of the vision and inertial sensors to ensure stable image acquisition in complicated road environment. Based on the research of robust and efficient unsupervised algorithms, the absolute depth information can be attained. Through the study of learning-based multi-task neural networks that integrate depth-fused semantic segmentation and object detection, we are able to acquire instant-level semantic results. Given the semantic and depth knowledge, we carry out studies of algorithms that distinguish dynamic from static targets as well as algorithms that fuse inertial navigation and vision sensor information, so as to improve the localization accuracy in high-dynamic scenes. And finally, by combining image, depth as well as semantic knowledge jointly, together with the research on movable probability prediction of static modeled objects, we aim at constructing a semantic point cloud mapping with high reusability. The research contents lay a critical foundation for autonomous decision of robots. The smooth development of the project will help to fill the research gap of the related fields at present stage. Furthermore, it can have a positive impact on improving the research level of our country’s automatic driving related technologies.

本项目将面向室外导航面临的高动态场景，基于仿生双目视觉技术开展鲁棒的室外即时定位与地图构建（SLAM）系统的研究。系统的各部分研究内容之间一脉相承：通过构建仿生双目运动控制模型，利用视觉与惯性传感器保障复杂道路环境下的视觉稳像采集；通过研究鲁棒而高效的无监督学习算法，获取场景的绝对深度信息；通过研究深度信息辅助的语义分割与物体检测多任务学习网络模型，得到实例级别的场景语义结果；基于语义与深度信息，研究动静目标的区分算法及视觉惯性信息融合算法，提升高动态场景下的系统定位精度；联合图像、深度与语义信息，加之对静态建模对象的移动概率预测研究，建立复用性较高的环境语义点云地图。本项目的研究内容，是机器人自主行为决策的重要感知基础。本项目的顺利开展，将填补现阶段相关领域的研究空白，对我国自动驾驶相关技术研究水平的提高，有着关键的促进作用。

本项目面向室外导航的高动态场景，基于仿生双目视觉技术开展鲁棒的室外即时定位与地图构建（SLAM）系统研究。系统的各部分研究内容紧密衔接：通过仿生双目运动控制模型，利用视觉与惯性传感器实现复杂道路环境下的视觉稳像采集；通过研究基于锚点引导的弱监督学习深度估计算法，获取场景的精准深度信息；通过研究深度信息辅助的语义分割与物体检测多任务学习网络模型，得到实例级别的场景语义结果；基于语义与深度信息，研究动静目标的区分算法及视觉惯性信息融合算法，提升高动态场景下的系统定位精度；联合图像、深度与语义信息，加之对静态建模对象的移动概率预测研究，建立复用性较高的环境语义点云地图。本项目最终建立了一套主动感知、定位精准且地图信息丰富合理的室外高动态SLAM系统。本项目构建了一套适用于室外高动态场景的SLAM及机器人导航系统，各项技术指标均达到提出的系统指标要求，其关键技术在机器人产业化项目中得到用户采用。受项目资助共发表机器人、计算机视觉期刊及会议论文10篇，申请专利10项。