同步定位与建图（Simultaneous Localization and Mapping，SLAM）是自主机器人在未知环境中依靠自载的传感器进行环境感知，实现运动估计和实时建图的一项技术。随着SLAM鲁棒时代的到来，激光雷达和视觉融合的SLAM系统能够发挥传感器信息互补的优势，提高系统在挑战性环境中的鲁棒性，同时随着人工智能的发展，利用人工智能等技术提高SLAM系统的信息感知和特征提取能力成为研究热点。本文针对室内结构化场景等挑战性环境，将深度学习技术引入到SLAM，在任务层面进行激光雷达和视觉的融合，构建一种同步激光-视觉SLAM的方案（VOLO），解决传统SLAM在长走廊等环境下适应能力不足的问题，本方案是一种鲁棒的松耦合形式融合的SLAM。主要研究内容如下：（1）设计了无监督学习的深度视觉里程计DS-SFMLearner，目的是提高模型在动态环境下的适应能力。DS-SFMLearner在SFMLearner的基础上采用光度不一致性作为视觉合成损失的权重，简化网络结构并解决原有的网络训练冲突问题。在KITTI数据集上的实验结果表明DS-SFMLearner位姿估计精度有显著提高，同时模型拥有对动态目标识别的能力，能够更好地应对动态场景下的位姿估计。（2）设计了VOLO面向融合的前端算法，目的是整合相机和激光雷达的优势，帮助ICP模型克服室内长廊等挑战性环境下的问题。VOLO前端算法利用DS-SFMLearner的位姿输出作为预估，ICP模型对位姿进行二次优化，通过实验表明前端融合算法不仅能够减少迭代计算时间从而提高效率，还能跳出局部极值点，提高位姿估计的精度。（3）设计了加权扫描上下文形式的回环检测算法，目的是利用点云的强度信息补充几何特征，解决在几何特征一致性较高环境下的回环误匹配问题。加权扫描上下文利用强度信息作为子空间高度特征的权重，增加了特征稀疏性，在KITTI数据集上的实验中提高了回环检测的精确率和召回率。为了评估上述构建的VOLO算法在实际室内环境中的表现，搭建了移动平台并进行真实室内长廊环境的数据采集，在自建数据集上对VOLO算法进行实验验证，实验表明VOLO算法相比纯LO算法能够提高定位和建图的性能，并提高效率。
Simultaneous Localization and Mapping (SLAM) is a technology that relies on self-loaded sensors for environment sensing, motion estimation, and real-time map building for autonomous robots in unknown environments. With the advent of the robustness-perception era for SLAM, SLAM systems with LiDAR and vision fusion can take advantage of complementary sensor information to improve the robustness of the system in challenging environments. Meanwhile, with the development of artificial intelligence, the use of artificial intelligence techniques to improve the information perception and feature extraction capability of SLAM is receiving attention. In this paper, we introduce deep learning technology to SLAM for challenging environments such as the indoor long corridor environment and build a simultaneous LiDAR-vision SLAM scheme to solve the problem of insufficient adaptability of traditional SLAM. The proposed SLAM is named VOLO, a robust SLAM with loosely coupled form of sensor fusion. The main research contents are as follows.(1) We design the unsupervised learning deep vision odometry, DS-SFMLearner, with the aim of improving the adaptive capability in dynamic environments.DS-SFMLearner adopts photometric inconsistency as the weight of visual synthesis loss based on SFMLearner to simplify the network structure and solve the problem of training conflicts in the original network. Experimental results on the KITTI dataset show that DS-SFMLearner has significantly improved the pose estimation accuracy and has the ability to recognize dynamic objects, which can better handle the pose estimation in dynamic scenarios.(2) We propose a fusion-oriented front-end algorithm for VOLO, which aims to integrate the advantages of vision and LiDAR to help ICP overcome problems in challenging environments such as indoor long corridors. The front-end algorithm uses the pose estimation of DS-SFMLearner as a proposal, and ICP algorithm is used for pose optimization. Experiments show that the fusion algorithm can reduce the iterative computation to improve the efficiency and jump out the local extremum points to improve the accuracy of the pose estimation.(3) We propose a loop closure detection algorithm in the form of weighted scan context. The purpose is to use the intensity information of the point cloud to complement the geometric features and solve the loop closure mis-matching problem in the environment with high geometric feature consistency. The weighted scan context utilizes the intensity information as the weight of the subspace height features, which increases the feature sparsity. The algorithm improves the accuracy and recall of loop closure detection in experiments on the KITTI dataset.To evaluate the performance of VOLO algorithm in real indoor environments, we build a mobile platform and use it to collect environmental data from indoor long corridors. We validate VOLO algorithm on the self-built dataset and the experiment shows that VOLO algorithm can improve the performance of localization and mapping and increase the efficiency compared to the pure LO algorithm.
