基于广义视觉的机器人直接定位与建图理论、方法研究

制约机器人视觉定位与地图构建性能的主要瓶颈有：1）传统相机具有有限的动态范围感知能力；2）基于稀疏图像特征的视觉定位方法对环境信息的利用率低。针对上述问题，我们借鉴多光谱视觉里程计和直接视觉SLAM的最新研究成果，提出基于广义视觉的直接定位与建图理论框架。首先，时空配准并融合多光谱仪、可见光相机、红外相机、激光雷达、TOF深度相机等传感器图像，通过金字塔像素模型建立广义的视觉信息表达，弥补第一个问题中关于环境信息感知不足的缺陷；其次，在广义视觉信息框架下，研究无须特征提取与匹配的直接定位与建图方法，解决第二个问题中关于图像信息使用率不高的问题。其中的难点是研究基于半致密金字塔像素模型的直接视觉里程计方法与视觉SLAM方法。最终目标是提高机器人定位与建图的精度，增强对白天与黑夜、室内与室外、人工与自然等不同环境的适应能力。

The main bottleneck of robotic localization and mapping using computer vision includes: 1) traditional cameras have limited dynamic range; 2) sparse features based localization method has a low utilization rate of environmental information. To solve these two problems, generalized vision based direct localization and mapping theoretical framework are proposed when we learn from the latest research results on Multispectral Visual Odometry (MVO) and Direct Visual SLAM. First, the data from multi-spectrometer, visible light camera, infrared camera, laser scanner and time of flight (TOF) depth camera are synchronized and aligned. A pyramid pixel model is used to represent the generalized visual information which is employed to solve the first problem on the lack of environmental information. Then, under the framework of generalized visual information, the direct localization and mapping method is studied without using the feature matching and tracking, which is aiming to solve the second problem on the issue of low usage of image information. Among these research contents, one of the difficulties is to study the semi-dense pyramid pixel model based direct VO and SLAM method. The ultimate goal is to improve the accuracy of localization and mapping for robot while enhancing the adaptive ability to the different environment such as day and night, indoor and outdoor, artificial and natural environments.