一种改进的地下车库弱纹理下智能汽车SLAM算法

地下车库中纯视觉的SLAM方法无法克服光线不足和弱特征纹理两大不利因素,为此论文提出一种基于VINS-Mono框架下改进的视觉惯导融合算法,把原算法中提取Harris角点的方法改进为提取灰度值陡变的像素点,并使用非线性优化方法在初始化阶段进行视觉位姿估计。后端采用滑动窗口的形式建立先验估计残差、IMU残差以及基于灰度值不变原理构建的视觉残差的联合残差模型,进一步提升了系统底层变量的优化效果从而提高算法的定位准确度。通过基于EuRoc数据集的仿真实验和地下车库实际场景的实车实验,验证了所提算法的鲁棒性和精确性。

An improved SLAM algorithm for smart cars under weak texture in underground garages

The pure visual SLAM (simultaneous localization and mapping) method in the underground garage cannot overcome two major disadvantages: insufficient light and weak feature texture. For this reason, a visual inertial navigation fusion algorithm that is improved based on the VINS-Mono framework is proposed in this paper. The method of extracting the Harris corner points from the original algorithm is improved to the method of extracting the pixels with sharp changes in gray value, and the nonlinear optimization method is applied to estimate the visual pose in the initialization stage. A sliding window is used at the back-end to establish a priori estimation residual, an IMU residual, and a joint residual model of visual residual that is constructed based on the principle of invariable gray value. The optimization effect of the system''s underlying variables is thus improved, so is the positioning accuracy of the algorithm. The robustness and accuracy of the proposed algorithm are verified through the EuRoc dataset-based simulation experiments and the real car experiments in the actual scene of the underground garage.