Distributed Sensor Networks for Scene Analysis in GPS Denied Environments

在许多国防和民用应用中，利用多个机载运动平台上的多个异质传感器对感兴趣区域进行主动和被动场景映射和探测是一个重要的问题。它在多目标跟踪、分类、导航、测绘等领域有着广泛的应用。用于此类问题的传感器模式可能包括电子支持措施(ESM)、多雷达(包括运动目标指示器(MTI)、红外线(提供高帧速率方位估计)和电光(EO)，包括激光雷达和基于相机的系统。每个传感器提供不同程度的精度、响应时间和性能，实际上，一个传感器获取的数据可以提高另一个传感器的精度。在GPS拒绝的环境中，从传感器阵列进行感知是具有挑战性的，因为传感器节点的位置是关键信息。此外，动态放置传感器阵列以通过传感器管理来增强场景分析，例如，通过进行无人驾驶飞行器(UAV)的指定运动，也取决于自定位数据。虽然航位推算技术可以提供帮助，但同时定位和跟踪(或映射)是关键的算法技术。传感器融合也是一个主要问题，即多个不同类型的传感器可能位于单一平台上，也可能分布在多个平台上，采用集中式或分布式数据融合，每种配置都提供了各自的挑战和机会。特别是，在分布式处理方法和集中式融合之间寻找最佳权衡，在分布式处理方法中，信息在传感器之间直接交换，集中式融合用于向操作员提供高级推理。对于传感器之间的信息交换，关键是要了解系统在存在干扰(从有源干扰到天气条件)时的能力，并纳入可以指示传感器性能下降的额外知识。尽管存在过多不同的传感器配置、隐含问题和潜在解决方案的组合，但每个传感器融合和管理问题都有几个共同的主题。其中包括1.了解如何以最佳方式量化和纳入辅助信息，如气象报告、目标机动模型、对轨迹的任何预期约束(例如民用飞行路径)和测量可靠性；2.了解不同传感器之间应直接交换哪些信息以增强最佳感知和检测，以及融合中心将如何纳入可用信息；3.了解实现场景分析和映射(包括目标跟踪、检测和分类)的有效算法。虽然很难将这些系统所需要的一切都结合在一起，但传感器网络的设计通常可以用多目标跟踪的概率图形模型来表示。这个博士项目将使用贝叶斯推理技术的最新进展，使用可扩展和灵活的消息传递框架，其中可以结合辅助信息，并隐含地实现数据关联和超参数估计。这个PHD将关注辅助信息的最佳表示，传感器超参数的表示，以及如何估计离散和连续混合变量的研究。该项目将建立在UDRC第3阶段WP1.2基础上，并对其进行补充。

Active and passive scene mapping and exploration of a region of interest using multiple- heterogeneous sensors on multiple airborne moving platforms is an important problem in many defence and civilian applications. It has many applications ranging from multi-target tracking (MTT), classification, navigation, surveying and mapping, and many others. Sensor modalities used in such problems may include electronic support measures (ESM), multiple-radars (including moving target indicators (MTI), infra-red (providing high-frame rate bearing estimates), and electro-optics (EO) including Lidar and camera-based systems. Each sensor provides varying degrees of accuracy, response time, and performance, and indeed the data acquired by one sensor can improve the accuracy of another.In GPS denied environments, sensing from an array of sensor arrays is challenging as the location of the sensor node is crucial information. Moreover, dynamic placement of a sensor array to enhance scene analysis through sensor management by, for example, making a designated movement of an uncrewed aerial vehicle (UAV), also depends on self-localisation data. Although dead-reckoning techniques can help, simultaneous localisation and tracking (or mapping) are key algorithmic techniques.Sensor fusion is also a major problem, whereby multiple heterogeneous sensors may be co-located on a single platform, or distributed across many platforms, with central or distributed data fusion, and with each configuration offering their own challenges and opportunities. In particular, finding the optimal trade-off between a distributed processing approach, in which information is exchanged directly between sensors, and a centralised fusion for delivering high-level inference to the operator. For information exchanged between sensors, it is crucial to understand the capability of the system in the presence of interference (from active jamming to weather conditions) and incorporating additional knowledge that can indicate the degradation in the sensor's performance.Although there is a plethora of different combinations of sensing configurations, implicit problems, and potential solutions to each scenario, there are several common-themes to each of these sensor fusion and management problems. These include1. understanding how to optimally quantify and incorporate auxiliary information, such as meteorology reports, models of target manoeuvres, any expected constraints on trajectories (for example civilian flight paths), and measurement reliability;2. understanding what information between heterogenous sensors should be exchangeddirectly to enhance optimal sensing and detection, and how the fusion centre will incorporate the available information;3. understanding efficient algorithms for enabling scene analysis and mapping (including targettracking, detection, and classification). Although difficult to incorporate everything that is desired in these systems, the design of the sensor network can generally be expressed in terms of probabilistic graphical models for multi-target tracking.This PhD project will use recent advances in Bayesian inference techniques using scalable and flexible message-passing framework, in which auxiliary information can be incorporated, and data association and hyper-parameter estimation is implicitly achieved. This PhD will be concerned with the best representation for auxiliary information, representations for sensor hyper-parameters, and investigation of how mixed discrete-and-continuous variables can be estimated. This project will build on and compliment the underpinning work in UDRC Phase 3 WP1.2.