基于Dirichlet过程非参贝叶斯学习的高斯箱粒子滤波快速SLAM算法

Fast SLAM Algorithm Based on the Non-parametric Bayesian Learning ofDirichlet Process for Gauss Box Particle Filtering

  • 摘要: 针对常规FastSLAM算法需要大量粒子创建地图以及粒子退化而导致计算复杂度高、难以提高估计精度等问题,提出了一种基于Dirichlet过程非参贝叶斯学习的高斯箱粒子滤波快速SLAM(同步定位与地图构建)算法.首先,改进了箱粒子滤波中以箱粒子为支撑集的均匀概率密度函数,采用高斯概率密度函数进行贝叶斯滤波,提高了估计的精度.在此基础上将Dirichlet过程非参贝叶斯学习应用于高斯箱粒子的重采样,既保证了有效箱粒子数,又能让箱粒子集中在高似然区域,降低了采样枯竭的影响.然后,利用基于Dirichlet过程非参贝叶斯学习的高斯箱粒子滤波进行机器人位姿估计,可有效降低地图创建所需的粒子数,并提高定位精度和实时性.进而采用无迹卡尔曼滤波更新地图特征,以提高地图创建的一致性.仿真结果和轮腿复合机器人实地实验结果验证了本文方法的可行性和有效性.

     

    Abstract: The traditional FastSLAM algorithm requires a large number of particles to build maps, and the particle degeneracy will lead to high computational complexity and difficulties in improving the estimation accuracy. For those problems, a fast SLAM (simultaneous localization and mapping) algorithm based on non-parametric Bayesian learning of Dirichlet process for Gauss box particle filter is proposed. Firstly, the uniform probability density function with box particles as the support set in box particle filter is improved, and the Gauss probability density function is used to carry out Bayesian filtering in order to improve the estimation accuracy. On this basis, non-parametric Bayesian learning of Dirichlet process is applied to resampling Gauss box particles, which not only ensures an effective number of box particles, but also enables the box particles to concentrate in the high likelihood region, and thus weakening the effect of sample exhaustion. Then, the non-parametric Bayesian learning of Dirichlet process for Gauss box particle filter is used to estimate the robot pose, which reduces the number of particles needed to build maps and improves the location accuracy and real-time performance. Further, the unscented Kalman filter is used to update the features in the map, which improves the consistency of map building. Simulation and experiment results on the wheel-legged robot verify the feasibility and effectiveness of the proposed method.

     

/

返回文章
返回