本文为希腊亚里士多德大学（作者：Michael Dille）的博士论文，共228页。

在许多救援、监视和科学应用中，广泛需要进行广域侦察和地形测量，无人驾驶飞行器（UAV）正日益受到欢迎。本文考虑的任务是使用一个或多个无人机来定位感兴趣的目标，提供连续的观测视角，并在目标因任何原因丢失时快速重新捕获跟踪。

对于一般的无人机来说，由于传感器视场小、无人机姿态估计不确定以及相对于环境规模的机动性有限，持续稳定跟踪目标是一项困难的任务，需要不断地处理观测数据，重新计算飞行路径或根据传感器数据以最好地找到目标或保持其在视野范围内。实现这一目标的现有策略只能提供对目标位置的糟糕估计，并且依赖于大量的启发式或计算密集型轨迹生成来进行追踪和搜索。

本文提出了观测不确定性和环境结构开发的精细表示方法，特别是以典型的城市道路网为例，来简化和更好地建模问题。对于主动跟踪的目标，通过为高不确定性观测设计的滤波器表示来证明极大地改进了位置估计，通过建模地形约束下的目标位置和运动空间缩减来提高跟踪性能。对于没有或只有大致已知先验位置的目标，需要进行初始搜索，同时考虑经典贝叶斯概率搜索（对于随机建模的运动目标）和新的道路网络覆盖策略（对于静止或缓慢移动的目标）。最后，该算法被扩展到搜索和局部找回道路网络中的逃逸规避目标，通过新的追踪策略进行映射，这些策略在抽象或地面领域中得到了很好的研究，但尚未在空中应用看到解决方案。

在使用广泛部署的飞行器进行的大量野外试验中，已经验证了本文设计的估计和跟踪方法，并且在实际仿真中使用类似参数化的飞行器模型和控制接口对其他部件进行了评估，为将演示算法直接应用于实践奠定了基础。

Across many rescue, surveillance, and scientific applications,there exists a broad need to perform wide-area reconnaissance and terrainsurveys, for which unmanned aerial vehicles (UAVs) are increasingly popular.This thesis considers the task of using one or more UAVs to locate an object ofinterest, provide continuous viewing, and rapidly re-acquire tracking should itbe lost for any reason.For both the common class of small field-launched UAVs considered aswell as larger UAVs, this is a difficult task due to a small available sensorfield of view, uncertain estimates of UAV pose, and limited maneuverabilityrelative to the scale of the environment, requiring constant processing ofobservations and recomputation of flight paths or sensor aiming to best findthe object or keep it in view. Existing strategies for accomplishing thisprovide poor estimates of the object’s location and rely on grossly heuristicor computationally intensive trajectory generation for both pursuit and search.

This thesis proposes careful representation of observationuncertainty and exploitation of environmental structure with particular focuson road networks typical of urban-like areas as means to simplify and better modelthe problem. For the case of actively tracked objects, greatly improvedlocation estimates are demonstrated through filter representations designed forhigh-uncertainty observations, as is increased pursuit performance by modelingterrain-constrained space reduction in object location and motion. Objectshaving no or only roughly known prior location require an initial search, forwhich both classical Bayesian probabilistic search (for stochastically-modeledmoving objects) and novel road network coverage strategies (for stationary orslow-moving objects) are considered. Finally, this is extended to search andlocal recapture of evasive adversaries in road networks through novel mappingsof pursuitevasion tactics that are well-studied in abstract or ground-baseddomains but have yet to see use in physical, particularly aerial, applications.Estimation and tracking aspects have been validated in extensivefield trials using widely-fielded air vehicles, and other components have been evaluatedin realistic simulation using similarly parameterized vehicle models andcontrol interfaces, laying the groundwork to directly apply the demonstratedalgorithms on real aircraft.

1. 引言

2. 问题研究

3. 搜索与跟踪的相关策略

4. 地理定位与跟踪

5. 覆盖与有效搜索

6. 搜索与重新捕获的保证

7. 结论

附录 小型无人机的鲁棒自动视觉跟踪

更多精彩文章请关注公众号：