Abstract:

Under unknown environment, Simultaneous Localization, Mapping and Object Tracking (SLAMOT) based on monocular vision needs sufficient parallax to meet the observability condition of object tracking. Focused on the uncertainty of target motion and the unknown of the system on target motion mode, a multi-mode filtering target tracking algorithm based on monocular suboptimal parallax was proposed. Firstly, the direction in which the target uncertain ellipsoid projection area changed the most was selected as the suboptimal parallax direction and was used as the robot parallax control direction. Then, multi-mode filtering algorithm was used to calculate the probability of different motion modes of the target, estimating the target state of different motion modes. Finally, the target state was estimated according to the probabilistic weighting of each motion mode. The simulation results show that the residual error of suboptimal disparity algorithm is 0.16 m when the parallax velocity is 0.3 m/s, meanwhile the residual means of heuristic algorithm, multimode filtering algorithm, traditional Extended Kalman Filter (EKF) algorithm are 0.25 m, 0.06 m and 0.16 m respectively. Besides, when the parallax speed is small, the proposed algorithm also can satisfy the observability condition of target tracking, having important engineering application value.

Key words: object tracking, Simultaneous Localization And Mapping (SLAM), multi-mode filtering, suboptimal parallax, control strategy

摘要：

在未知环境下基于单目视觉的机器人同时定位、地图构建和目标追踪的耦合问题（SLAMOT）中，需要足够的视差才能满足目标跟踪的可观性条件。同时，针对目标运动的不确定性以及系统对于目标运动方式的未知性，提出一种基于次优视差的多模滤波目标跟踪算法。首先，采用目标不确定性椭球投影面积变化最大的方向为次优视差方向，并将其作为机器人视差控制方向；然后，采用多模滤波算法计算目标各种运动方式的概率；其次，对各运动方式的目标状态进行估计，最后根据各运动方式的概率加权估计出目标状态。另外，考虑到工程应用中应减小能耗，因此，在满足目标跟踪要求的条件下，降低视差速度。仿真实验表明：视差速度为0.3 m/s时，次优视差算法的残差均值为0.16 m，而启发式算法、多模滤波算法、传统扩展卡尔曼滤波（EKF）算法的残差均值为0.25 m、0.06 m和0.16 m。在视差速度较低时，所提算法也能满足目标跟踪的可观性条件，具有较强的工程应用价值。

关键词: 目标跟踪, 同时定位与地图构建, 多模滤波, 次优视差, 控制策略