Single and Cooperative Imaging based GPS Denied Localisation Solution for Autonomous Platforms.

为了解决基于摄像机的多目标跟踪系统的可靠和高效的基本问题，采用视觉里程计（VO）和被动热成像（TI）。目的是产生一个明显更方便的解决方案，能够在网络中的传感器范围之间跟踪和分类目标。对解决方案的主要干扰来自云和其他遮挡物，其他问题包括不同的照明水平；如果平台希望保持隐身，这一点将尤为重要。视觉里程计已广泛应用于机器人应用中，通过定期拍摄的图像来估计物体的运动。由于Plenoptic技术的进步，视觉里程计和Plenoptic技术的结合很可能胜过类似的算法，如同步定位和地图绘制（SLAM）。由于VO侧重于相对位置而不是绘制地形，因此有可能通过使用深度映射来增强基线解决方案。因此，这提供了创建传感器基础网络的可能性，该网络可以识别多个目标并跟踪其在整个网络中的传播，为有效的实时多目标跟踪解决方案提供骨干。使用被动热成像技术可以解决光照水平变化带来的挑战，因为TI检测的是物体的发射度，而不是其反射，因此可以在夜间使用。使用TI是有问题的，因为它会使解决方案容易出现低空间分辨率、历史效应、温度变化和低信噪比。这是一个令人兴奋的挑战，是为这些特殊类型的传感器开发新的图像处理方法的绝佳机会。使用深度卷积神经网络和深度Q学习有可能解决可靠和有效的目标识别问题。近年来，深度学习已经发展到可以实现实时目标检测和分类的地步——即使是在相对平淡的机器上。这意味着传感器网络中深度卷积神经网络和LSTM技术的结合为多目标跟踪系统提供了一种可能的解决方案。

To employ the use of Visual Odometry (VO) and the Passive Thermal Imagery (TI) in order to tackle the fundamental problem of reliable and efficient camera based multi-target tracking systems. The aim is to produce a significantly more convenient solution that is capable of tracking and classifying targets between the ranges of sensors in the network. The main interference to the solution would arise from cloud and other obscurants, other problem would include varying illumination levels; this would be particularly significant should platforms wish to remain stealthy. Visual Odometry has been used in wide range of robotic applications that estimate the motion of an object from images taken at regular intervals. Due to advances in Plenoptic technology, it is likely that unison of visual Odometry and Plenoptic technology would outperform similar algorithms such as Simultaneous Localisation and Mapping (SLAM). As VO focuses on relative position(s) rather than mapping the terrain, there exists a possibility to enhance a baseline solution by the use of depth mapping. This therefore provides the possibility to create a sensor base network which can identify multiple targets and track their propagation throughout the network providing the backbone to an effective real time multi-target tracking solution. The challenge of varying illumination levels can be addressed by the use of passive thermal imagery - due to the fact TI detects an objects emittance rather than its reflection and so can be used at night. The use of TI is problematic as it would make the solution prone to low spatial resolution, history effects, variations in temperature, and low signal-to-noise ratios. This is an exciting challenge an excellent opportunity to develop new image processing methods for these particular types of sensors.The use of Deep Convolutional Neural Networks and Deep Q Learning has the potential to tackle the problem of reliable and efficient target recognition. In recent years deep learning has advanced to the point that real time object detection and classification have become feasible - even on relatively lacklustre machines. This means the combination of Deep Convolutional Neural Networks and LSTM techniques employed on the sensor network provides a possible solution to a multi-target tracking system.