结合道路结构化特征的语义SLAM算法

Semantic SLAM algorithm combined with road structured features

摘要:

视觉SLAM（simultaneous localization and mapping）是智能车辆领域的研究热点,在包含运动目标干扰或近景特征不显著的场景中,容易产生帧间位姿估计结果精度不足或失效问题.为此,本文提出一种结合场景语义信息和路面结构化特征的SLAM算法.首先,针对上述特殊场景中运动目标干扰的情况,设计带有改进金字塔池化模块的语义分割神经网络,得到图像中各像素对应的目标类别,作为剔除运动像素点的依据,从而避免运动点参与特征匹配导致的位姿计算准确性下降问题；然后,针对有效近景特征点不足的情况,基于V视差算法确定图像中的道路平面区域并拟合出精确的视差方程,以计算路面上像素点的精确视差值,并提出一种基于路面结构化特征（车道线、马路边界、路面交通标记等）的位姿计算方法；最后通过场景实验得出,本文提出的改进算法计算结果的绝对轨迹误差小于原算法.证明该方法能够在存在运动目标干扰或缺乏近景特征的场景中具有较高的位姿估计精度,建立了有效的包含语义信息的稠密点云地图,具有良好的环境适应性.

Abstract:

Vision-based simultaneous localization and mapping (SLAM) is a research hotspot in the field of intelligent driving. However, for the scenes that contain moving targets or inconspicuous close-range features, it is easy to cause ineffective or inaccurate pose estimation between frames. To solve this problem, this paper proposes a SLAM algorithm based on road structured features and scene semantic information. First, for the problem of target moving, a semantic segmentation neural network with an improved pyramid pooling module was designed to obtain the target category corresponding to each pixel in the image. The segmentation results were taken as basis for the elimination of moving points, which avoids the problem of low accuracy of pose calculation caused by moving points participating in feature matching. Then, in view of the lack of effective feature points, the road area in the image was determined based on v-disparity algorithm, and disparity function was obtained to calculate the accurate disparity value of the pixels on the road. Furthermore, a pose calculation method based on road structured features (e.g., lane lines, road boundaries, pavement traffic markings) was proposed. Finally, scene experiments were carried out and results show that the absolute trajectory error of the improved algorithm proposed in this paper was smaller than that of the original algorithm, which proves that the proposed method has higher pose estimation accuracy in scenes with moving targets or inconspicuous close-range features. In addition, an effective dense point cloud map containing semantic information was established, which has good environmental adaptability.