基于协同演化的动态多目标追踪

COOPERATIVE COEVOLUTION FORDYNAMIC MULTI-TARGET TRACKING

赖卓君

11849309

硕士

电子与通信工程领域工程

史玉回

2020-05-30

2020-07-20

哈尔滨工业大学

深圳

动态多目标追踪涵盖了机器人搜索、跟随、避障等机器人领域的核心应用，具有广泛的应用前景。动态多目标追踪问题属于追逃问题，追逃问题是一个经典问题，受到了众多研究领域的关注，例如博弈论、路径规划、强化学习和多机器人系统等领域。相关研究有很多，但是其中大多数研究都是倾向于将追逃问题作为一个应用，来验证其理论的有效性，并不适合实例化为真实的机器人使用。所以本文的目的是设计一个更符合真实应用场景的动态多目标追踪系统，并在物理仿真器Gazebo上做实验模拟，可以应用于巡逻、跟拍、监护、搜救和社交导航等场景。在本文中有两个机器人群体，其中一群机器人作随机运动，称为逃跑机器人；另外一群机器人可以相互通信、具备局部感知能力，它们的任务是找到地图中的逃跑机器人，并均匀地包围且跟随在其周围，期间不与障碍物或者其他机器人发生碰撞，为追捕机器人。本文将动态多目标追踪问题建模成一个动态优化问题，优化的对象是追捕机器人的目标点，优化的目标有八个：避障、避碰、探索未被访问的区域、减小角速度、接近全局轨迹子目标点、和逃跑机器人保持距离、均匀包围逃跑机器人和相机方向正对目标机器人。采用协同演化头脑风暴优化算法（cooperative co-evolutionary brain storm optimization, CCBSO）将含有N个追捕机器人的追踪问题分解成N个子问题，即由同时优化N个目标点的问题转变为N个目标点的优化问题。每个子问题都由单独的子种群解决，子种群中每个个体的适应值评估依赖于子种群间的协作，最后的完整解由每个子种群优化的结果组合而成。本文构建了动态多目标追踪系统，主要包括两个模块：感知模块和导航模块。感知模块主要用于检测障碍物与逃跑机器人；导航模块主要用于决定机器人的状态以及控制机器人运动，包括全局路径规划器、行为规划器以及局部路径规划器。其中，全局路径规划器用于动态地规划追捕机器人到每个已捕捉逃跑机器人的全局轨迹；行为规划器决定了追捕机器人当前的任务，任务分为搜索、追踪、监视和跟随；局部路径规划器则采用CCBSO算法优化追捕机器人的目标点，再由目标点反推出其速度控制信号控制机器人运动。实验表明，对于速度较慢的逃跑机器人来说，追捕机器人能够有效地进行捕捉和包围。

Dynamic multi-target tracking covers the core applications in robotics such as searching, following and obstacle avoidance, which has broad application prospects. The dynamic multi-target tracking problem belongs to the pursuit-evasion problem, which is a classic problem. It has attracted the attention of many research fields, such as game theory, path planning, reinforcement learning, and multi-robot systems. There are many related studies, but most of them tend to treat the pursuit-evasion problem as an application to verify the validity of their theory, which is not suitable for instantiation into real robots. So the purpose of this thesis is to design a dynamic multi-target tracking system that is more suitable for real application scenarios, and make an experimental simulation on thephysical simulator Gazebo. There are two groups of robots in this thesis. A group of robots make random movements, called evaders; another group of robots can communicate with each other, and their task is to find the evaders in the map, and evenly surroundand follow around them. They do not collide with obstacles or other robots during this period, called pursuers.This thesis models the dynamic multi-target tracking problem as a dynamic opti-mization problem. The decision vector is composed of the target point of pursuers, and there are eight optimization objectives: obstacle avoidance, collision avoidance, exploring unvisited areas, reducing angular velocity, approaching global sub-target point, keeping distance from the escape robot, evenly surrounding the escape robot and the camera facing the target robot. Using cooperative co-evolutionary brain storm optimization algorithm (CCBSO) to decompose the tracking problem with N pursuers into N subproblems, whichis transforming the optimization of N target points problem to N problems of optimizing target point. Each subproblem is solved by a separate subpopulation, the evaluation of the fitness value of each individual in a subpopulation depends on the collaboration among subpopulations, and the final complete solution is composed of the results of the optimization of each subpopulation. This thesis builds a dynamic multi-target tracking system mainly includes two modules, a perception module and a navigation module. The perception module is mainly used to detect obstacles and evaders; the navigation moduleis mainly used to determine the robot’s behavior and control the robot’s motion, including global path planner, behavior planner, local path planner. The global path planner is used to dynamically plan the global path of the pursuer to each captured evader; the behaviorplanner determines the current task of pursuers. The tasks are divided into searching, tracking, monitoring and following; the local path planner uses CCBSO to optimize thetarget point of the pursuer, and then use the target point to calculate its velocity controlsignal to control the robot’s movement. Experiments show that, for the evader with a lowspeed, the pursuer can effectively capture and surround it.

追逃问题 协同演化 头脑风暴优化算法 社交导航

pursuit-evasion cooperative coevolution brain storm optimization social-aware navigation

中文

联合培养

南方科技大学

赖卓君. 基于协同演化的动态多目标追踪[D]. 深圳. 哈尔滨工业大学,2020.