基于卷积神经网络的移动机器人自适应光照增强单目视觉SLAM算法

Adaptive Illumination Enhanced Monocular Vision SLAM Algorithm for Mobile Robots Based on Convolutional Neural Networks

摘要:

移动机器人视觉SLAM技术能够在一定条件下实时估计自身在环境中的位置，并构建和更新环境稀疏或稠密三维地图，这些信息可以帮助机器人提高对未知复杂环境的准确感知和适应能力，以执行更复杂的任务。但使用相机作为传感器的视觉SLAM在定位和建图的精度和稳定性方面在很大程度上依赖于采集到的图像质量，在弱光照环境中，现有的视觉SLAM算法难以有效地工作。针对视觉SLAM在弱光照环境中定位精度降低和跟踪丢失的问题。本文提出了一种适应弱光照环境的RLMV-SLAM算法，该算法使用一个轻量化的神经网络对输入图像进行预处理，增强其亮度、对比度、色彩和去噪，同时，该算法使用地图点补充策略、Sparse BA和一种实时增量闭环检测方法提高了定位和建图精度和鲁棒性。在公开数据集和自采数据集上对该算法进行了实验验证，并与其他主流视觉SLAM方法进行了对比，结果表明本文提出的方法将弱光照环境中有效跟踪时长提升30%以上，且在公开数据集上估计位姿的误差也有明显降低，证明了所提算法的有效性，为弱光照环境中同步定位和建图提供了一定参考。

Abstract:

The visual SLAM technology of mobile robots can estimate their position in the environment in real time under certain conditions, and build and update sparse or dense 3D maps of the environment. This information can help robots improve their accurate perception and adaptability to unknown complex environments, and perform more complex tasks. However, the accuracy and stability of localization and mapping of visual SLAM using cameras as sensors largely depend on the quality of the collected images. In low-light environments, existing visual SLAM algorithms have difficulty working effectively. In response to the problems of reduced positioning accuracy and lost tracking faced by visual SLAM in low-light environments, a visual SLAM algorithm suitable for low-light environments, RLMV-SLAM was proposed. This algorithm used a lightweight neural network to preprocess the input images, enhancing their brightness, contrast, color, and denoising. At the same time, the algorithm applied a map point supplement strategy, Sparse BA, and a real-time incremental loop closure detection method to improve the accuracy and robustness of localization and mapping. The research experimentally verified this algorithm on public datasets and self-collected datasets, and compared it with other mainstream visual SLAM methods. The results showed that the method proposed can increase the effective tracking time in low-light environments by more than 30% and significantly reduce the pose error of pose estimation on public datasets, proving the effectiveness of the proposed algorithm and providing a reference for simultaneous localization and mapping in low-light environments.