Defense

Robust detection strategies with machine learning and compressive sensing

The detection of targets in radar signals is a fundamental process. In areas where huge volumes of data are available to train neural networks (NN), machine learning (ML) has been able to achieve tremendous success, for instance in pattern, image, and speech recognition. The combination of compressive sensing (CS) and ML can solve the problem of missing data volumes, making these methods applicable in this field as well.

Typical neural network topology for the detection of structures.
© Fraunhofer FHR
Typical neural network topology for the detection of structures.
The k plane of a vector approximate message passing (VAMP) networks which allows for the reproduction of iterative CS algorithms.
© Fraunhofer FHR
The k plane of a vector approximate message passing (VAMP) networks which allows for the reproduction of iterative CS algorithms.
For interferometric SAR images of urban areas, there are several point scatterers in a distance azimuth cell. These can be detected and presented individually with the combined ML/CS method. The image shows the first two detections with their amplitudes and heights.
© Fraunhofer FHR
For interferometric SAR images of urban areas, there are several point scatterers in a distance azimuth cell. These can be detected and presented individually with the combined ML/CS method. The image shows the first two detections with their amplitudes and heights.

The main task of radar systems is to detect moving objects and, if possible, to determine the associated parameters, such as distance and speed. Conditioned by the selected frequency range of the systems and their active emission of coded signals, this capacity is independent of weather influences and the times of the day. Interfering signals such as multiple reflections, noise, or even explicit interference attempts dramatically complicate the detection of weak targets. For moving objects, there is an additional challenge, as the estimation of the parameters is made very difficult by the non-linear movement in relation to the radar system.

In the past years, one research focus was the improvement of the detection using additional external information. This information was used to enrich the actual measurement with additional knowledge to possibly also complement the measurement with information that is not measured, for instance the information that there are only a few moving objects in the observation space. This knowledge is the basis of the »sparse signal reconstruction« or also the »compressive sensing« theory. Fraunhofer FHR already recognized this advantage quite some time ago and thus created the International Workshop on Compressed Sensing applied to Radar, Multimodal Sensing, and Imaging (CoSeRa) in 2012 and intensely supported the workshop in the following years in a similar manner. In 2018, the fifth CoSeRa workshop was held in cooperation with the University of Siegen.

A known disadvantage of the CS methods compared to the classical methods is the high computing power required to reconstruct the observed scene from the measurement data before further signal processing steps can occur. In contrast to CS, classical methods, which filter the measurement data in contrast to CS in such a way that only moving targets remain.

Parallel to CS, the machine/deep learning (ML/DL) research area has led to a renaissance of neural networks in the past years, with considerable success in many commercial applications such as automatic face recognition, speech recognition, machine translation, driverless cars, and robotics. Important components for the success of ML/DL methods are, on one hand, the availability of large volumes of data necessary to train and condition the complex neural networks and, on the other hand, the further development of computing power as well as the current capacity to create non-linear models and adapt them to the existing data. When the conditioned neural networks are available, the ML/DL methods have the potential to provide computationally efficient approaches to improve the target detection, tracking, and classification in radar with an improved resolution.

An active field of research in this context is the integration of compressive sensing and machine learning for radar applications, thus combining the advantages of both methods. For example, iterations of many CS algorithms for the restoration of sparse signals have the structure of neural network layers. Therefore, numerically efficient ML models such as deep neural networks can be used to replace the computationally intensive iterations with fixed depth feedforward networks derived from training data. Furthermore, the learning process allows for the extraction of better dictionaries / transfer functions for the presentation of the observed scene based on the training data. CS-based generative models of targets and interfering signals can be used to create extensive training sets, which are necessary for the ML/DL algorithms. Thus, gaps in the measurement data can be closed by online predictions from a »compressed« database to improve the generalization. This CS-supported training strategy will significantly expand the range of applications in which ML/DL can be used successfully as the availability of large radar data sets is not always guaranteed in the military field.