面向动态场景复合深度学习与并行计算的DG-SLAM算法

DG-SLAM algorithm for dynamic scene compound deep learning and parallel computing

针对现有的同时定位与建图(SLAM)算法实时性不高和在动态环境中定位精度会大幅降低的缺点,提出了一种复合深度学习与并行计算的DG-SLAM算法。采用基于深度学习的目标检测算法检测出行驶环境中的动态物体,在ORB-SLAM2图像帧间匹配前剔除动态物体特征点,降低动态物体对SLAM定位精度的影响;在ORB-SLAM2跟踪局部地图中采用三维空间下内部点的判别方法区分内点和外点,建立GPU并行计算模型以高效搜索局部地图点;利用Saturated核函数作用于重投影误差项的二范数平方和,确保局部地图优化位姿时重投影误差的并行计算。在KITII数据集上进行了算法验证,结果表明,DG-SLAM具有较高跟踪精度,且平均计算效率相同情况下对比ORB-SLAM2高3.4倍以上,超过85帧/s,可实现自动驾驶车辆在动态环境下SLAM系统的稳定运行。

In view of the disadvantages of the existing simultaneous localization and mapping(SLAM)algorithm,which has low real-time performance and the positioning accuracy is greatly reduced in dynamic environment,a DG-SLAM algorithm based on deep learning and GPU parallel computing was proposed.The deep learning-based object detection algorithm was introduced to detect dynamic objects in the driving environment,and the feature points of dynamic objects were removed before the matching of image frames,so as to eliminate the impact of mismatching dynamic objects on the positioning accuracy of SLAM system.In the tracking of local maps of ORB-SLAM2,the discriminant method of 3D interior points was used to distinguish the inner points and outer points,and the GPU parallel computing model was established to efficiently search the local map points.The Saturated kernel function was used to minimize the reprojection error of the two norm terms to ensure the parallel calculation of the reprojection error when the local map was optimizing.The algorithm was verified on the KITII dataset,DG-SLAM has high tracking accuracy,and the average calculation efficiency was more than 3.4 times faster than that of ORBSLAM2 system under the same conditions,more than 85 frames per second,which could realize efficient and high precision SLAM in dynamic scene.