【计算机科学】【2018.12】基于POMDP的水下机器人多视点多目标分类路径规划方法

本文为美国康奈尔大学（作者：Quanxing Lu）的硕士论文，共39页。

本文提出了一种利用水下机器人成像传感器在最短时间内对多个感兴趣目标进行分类的方法。总体目标是通过依次求解单个目标分类问题和全局目标排序问题来实现的。首先，基于POMDP框架提出了一种基于深度卷积神经网络和支持向量机的多视点单目标分类算法。分类算法允许水下机器人在目标附近自适应地选择下一个配置状态，以最大限度地提高分类可信度。其次，利用旅行商算法生成全局目标访问订单。以一个装有侧扫声纳的无人水下机器人为例，仿真结果验证了该算法的有效性，并证明了该算法在基于多视点的多目标分类中能够找到明显较短的路径。

This thesis presents an approach of classifyingmultiple targets of interest in minimum time with satisfactory confidence by animaging sensor on an underwater robot. The overall goal is achieved bysequentially solving a single target classification problem and a global targetordering problem. First, a multi-view single-target classification algorithm isdeveloped based on the POMDP framework, which incorporates a deep convolutionalneural network and a support vector machine as the observation model. Theclassification algorithm allows the underwater robot to adaptively select itsnext configuration state near the target of interest in order to maximize theincrease of classification confidence. Next, a traveling salesman algorithm isused to generate the global target visiting order. Simulation results of anunmanned underwater vehicle equipped with a side-scan sonar validate theeffectiveness of the proposed algorithm and demonstrates the ability to findsignificantly shorter path for multi-view based multi-target classification.

1. 引言
2. 问题描述
3. 通用的POMDP公式
4. 关于图像模板的问题
5. 基于POMDP的水下机器人路径规划
6. 数值分析的结果
7. 结论与展望

更多精彩文章请关注公众号：