基于逆深度滤波的双目折反射全景相机动态SLAM系统

Dynamic SLAM of binocular catadioptric panoramic camera based on inverse depth filter

为了在复杂的动态环境下能够准确地估计折反射全景相机的位姿,提出一种基于逆深度滤波的双目折反射全景相机动态SLAM系统。分析了双目折反射全景相机系统,在球面进行三角化计算出空间点的深度信息。采用一种基于贝叶斯滤波的逆深度滤波器,对动态地图点进行剔除,只使用静态地图点对折反射全景相机进行位姿估计,可以在动态环境下得到精确的相机位姿估计结果。分别在低动态环境和高动态环境下进行了实验,实验结果表明,在高动态环境中,传统的SLAM系统计算的轨迹出现漂移,而基于逆深度滤波的双目折反射全景相机动态SLAM系统始终可以稳定运行。所提方法在两段高动态环境下的绝对轨迹误差的均方根误差分别为0.52 m和0.56 m,精度比传统的全景SLAM系统提升了96.35%。

To achieve accurate pose estimation of a catadioptric panoramic camera in complex dynamic environments,this study proposed a binocular catadioptric panoramic camera simultaneous localization and mapping(SLAM)system for dynamic environments based on an inverse depth filter. The binocular catadioptric panoramic camera system was analyzed,and the depth information of space points was calculated through the process of triangulation on a sphere. An inverse depth filter based on Bayesian filtering was used to eliminate dynamic map points,whereby only static map points were then used to estimate the poses of the camera. Accurate pose estimation results were obtained in dynamic environments. More specifically,experiments were conducted in both low and high dynamic environments,where results showed that when the trajectory calculated by traditional SLAM system drifted,the dynamic SLAM system based on inverse depth filtering was always executed in a stable manner. The root mean square errors of the absolute trajectory error of our method in the two groups of highly dynamic environments are 0. 52 m and 0. 56 m,respectively. The accuracy is up to 96. 35% higher than that of the traditional panoramic SLAM system.