现有的视觉SLAM方法大多依赖于静态场景以及环境光照稳定的假设，从而限制了视觉SLAM方法在室外动态环境下的应用。本项目以语义图模型理论为基础，研究室外动态环境的视觉SLAM问题。首先，结合图像语义分割技术构建一种动态混合地图节点语义图模型及对应的联合概率分布表达式；其次，基于深度卷积网络设计可学习图像特征点提取前端架构及其在线学习方法；然后，结合语义分割以及距离变换技术设计语义增强代价函数并将其与语义图模型对应的代价函数相结合，建立基于增量优化理论及EM算法的在线最大似然解求解方法；最后，为了训练得到可学习图像特征点提取前端的基础模型并且能够定量评估系统性能，研究基于真实感物理引擎自动构建大规模视觉SLAM基准数据库。项目将语义图模型理论及其在线求解方法应用于室外动态环境的视觉SLAM问题研究，有望为解决视觉SLAM在包含光照变化的实际动态场景的应用提供新的理论和技术支持。

Traditional visual SLAM heavily relies on the assumption that the environment and the illumination are static, which prohibits it from applications in outdoor dynamic environments. Based on the theory of semantic graphical model, the visual SLAM problem in outdoor dynamic environments will be studied in this project. Firstly, a semantic graphical model which contains dynamic object velocity nodes and mixed map nodes together with its corresponding joint probability distribution will be proposed with the help of semantic image segmentation technology. Secondly, a learnable image feature points extraction front-end will be designed based on deep convolutional neural network. The online learning algorithm of the front-end will also be developed in this project. Then, semantic enhanced cost function based on semantic segmentation and distance transformation will be designed and combined with the cost function which corresponds to the semantic graphical model. The online MAP solving method will be proposed based on incremental optimization theory and EM algorithm. Finally, for the sake of training the basic model of the learnable image feature points extraction front-end and evaluating the performances of the proposed systems quantitatively, we plan to utilize high fidelity physical game engine in order to automatically construct a large scale visual SLAM benchmark dataset. This project applies the theory of semantic graphical model and its corresponding online learning method to the problem of visual SLAM in outdoor dynamic environments. It is hoped to provide novel theories and technologies for visual SLAM system aimed for real dynamic environments which include illumination variations.