Machine Learning for Target Detection and Classification Using Multi-Modal Airborne Sensor Data

多年来，机器学习(ML)算法已经成功地应用于视觉图像数据库中，以识别场景中的对象。工作在射频(RF)的传感器大多产生低分辨率数据，其中检测到的信号更抽象，需要算法处理才能将信息呈现给人类操作员。射频传感器可以主动工作(检测自身发射的物体的反射信号)和被动工作(截获物体的发射和反射)。目标的特性可能取决于射频发射的反射和截获，后者可能处于不同的频率，并且通常与环境有关。久负盛名的经典检测方法往往通过去除所有看起来不像感兴趣的信号来发挥作用，这可能会在这个过程中丢弃有价值的信息和背景。该过程通常对每种类型的数据分别执行，并且仅合并处理后的输出。多传感器数据流的实时协同处理在分析和计算上都是具有挑战性的。这些数据的融合通常直到个别经典检测过程之后才会发生。然而，ML技术可能能够学习从场景的传感器数据中提取有益的特征，以检测所有感兴趣的对象，并在处理中比通常可能的更早提供对象类型的分类，并且与经典方法相比具有更好的置信度。本博士将寻求建立是否可以使用信号处理链早期使用的多模式机载传感器数据来更早地检测和分类对象，从而提高检测和分类性能。

Machine learning (ML) algorithms have been applied successfully for many years to databases of visual imagery for the recognition of objects in a scene. Sensors working at radio frequency (RF) mostly produce low-resolution data where the signals detected are more abstract and require algorithmic processing to present the information to the human operator. RF sensors can operate actively (detecting reflected signals from objects illuminated by their own transmission) and passively (intercepting emissions and reflections from objects). The characteristics of a target can depend on both reflections and interceptions of RF emissions, which are likely to be at different frequencies and are often context dependent. Long-established classical detection methods tend to work by removing everything that does not look like the signal of interest, which may throw away valuable information and context in the process. This process is typically performed on each type of data separately and only the processed outputs are combined. Real-time co-processing of multiple sensor data streams is analytically and computationally challenging. Fusion of this data usually does not occur until after the individual classical detection processes. However, ML techniques may be able to learn to extract the beneficial features from sensor data of the scene to detect all objects of interest and provide classification of object types earlier in the processing than is usually possible, and with improved confidence compared to classical methods.This PhD will seek to establish whether multi-modal airborne sensor data, used early in the signal processing chain, can be used to detect and classify objects earlier in the processing chain leading to improved detection and classification performance.