Automatic Target Recognition: Fundamentals and Applications

By Fouad Sabry

What Is Automatic Target Recognition

The capacity of an algorithm or device to detect targets or other objects based on data acquired from sensors is referred to as automatic target recognition, abbreviated as ATR.

How You Will Benefit

(I) Insights, and validations about the following topics:

Chapter 1: Automatic target recognition

Chapter 2: Computer vision

Chapter 3: Radar

Chapter 4: Synthetic-aperture radar

Chapter 5: Beamforming

Chapter 6: Pulse-Doppler radar

Chapter 7: Inverse synthetic-aperture radar

Chapter 8: Radar signal characteristics

Chapter 9: Time delay neural network

Chapter 10: Track algorithm

(II) Answering the public top questions about automatic target recognition.

(III) Real world examples for the usage of automatic target recognition in many fields.

(IV) 17 appendices to explain, briefly, 266 emerging technologies in each industry to have 360-degree full understanding of automatic target recognition' technologies.

Who This Book Is For

Professionals, undergraduate and graduate students, enthusiasts, hobbyists, and those who want to go beyond basic knowledge or information for any kind of automatic target recognition.

• Language: English
• Publisher: One Billion Knowledgeable
• Release date: Jul 4, 2023

Book preview

Chapter 1: Automatic target recognition

The term automatic target recognition (ATR) refers to the capability of an algorithm or device to identify a target or other object from sensor data.

In the early days of radar, operators had to listen to audio representations of received signals and use their training to determine what kind of target was being illuminated. Despite the accomplishments of these human experts, automated methods have been and are being developed to improve classification accuracy and speed. Animals, humans, and cluttered vegetation are just some of the biological targets that can be identified with ATR technology. Application areas range from object recognition in the field to reducing noise from birds on Doppler weather radar.

Possible military uses include an easy-to-implement identifier like an IFF transponder, as well as more complex systems like UAVs and cruise missiles. The potential uses of ATR in the home arena are attracting more and more attention. A variety of applications, from automated vehicles to safety systems that can detect objects or people on a subway track, to border security, have benefited from ATR studies.

Almost as long as radar has been around, targets have been recognized. Radar operators would use the audible representation of the reflected signal to identify enemy bombers and fighters (see Radar in World War II).

For a long time, operators would listen to the baseband signal to identify targets. Trained radar operators can use this signal to determine the type of vehicle being used to illuminate the target, the size of the target, and possibly even the presence of biological targets. But there are many restrictions on this method. There is a high probability of error due to the human decision component, the need for the operator to be trained for what each target will sound like, and the possibility that the target will no longer be audible if it is traveling at high speeds. This concept of aural representation of the signal, however, did lay the groundwork for automated target classification. Features of the baseband signal that have been used in other audio applications, such as speech recognition, have been incorporated into a number of classification schemes that have been developed.

The range of an object can be calculated using radar by timing how long it takes for the signal to return from the target that the signal illuminates. The Doppler effect describes the modification of frequency caused by the motion of such an object. A frequency shift can be caused by vibration or rotation of an object in addition to the translational motion of the whole object. The Doppler-shifted signal will become modulated if this occurs. The micro-Doppler effect refers to the additional Doppler effect responsible for the signal modulation. In order to create algorithms for ATR, this modulation can have a recognizable signature. When the target is moving, the micro-Doppler effect will cause a time- and frequency-varying signal.

Since the Fourier transform does not take time into account, analyzing this signal with a Fourier transform is insufficient. The short-time Fourier transform is the most straightforward approach to obtaining a frequency-time function (STFT). The frequency and time domains can be represented simultaneously using more robust methods like the Gabor transform or the Wigner distribution function (WVD). However, frequency resolution and time resolution will always be compromised in these approaches.

After this spectral data has been extracted, it can be compared to a database containing information about the targets the system will identify in order to determine what the illuminated target actually is. In order to determine which target in the library best fits the model built using the received signal, a statistical estimation method such as maximum likelihood (ML), majority voting (MV), or maximum a posteriori (MAP) is used.

Automated target recognition systems that use audio features from speech recognition to determine a target's identity have been the subject of research. Some examples of these coefficients are:

LPC Coefficients, or Linear Predictive Codes

Coefficients of linear prediction and coding in the cepstral spectrum

Cepstral coefficients based on mel frequencies (MFCC).

These coefficients are derived from a processed baseband signal, and a statistical method is then used to determine which target in the database is most similar to the derived coefficients. The system and use case must be taken into account when deciding which features and decision scheme to implement.

Target classification features are not restricted to coefficients motivated by human speech. ATR can be achieved using numerous feature sets and various detection methods.

Developing a training database is necessary for automating target detection. Typically, the ATR algorithm is fed experimental data collected after the target has been determined.

The flowchart represents one type of detection algorithm. This technique takes M blocks of data, models them using a Gaussian mixture model, and then uses the extracted features (such as LPC coefficients or MFCC) to draw conclusions (GMM). After fitting the data to a model, conditional probabilities are calculated for each target in the training set. Here, we have M data blocks to examine. This will generate M individual probabilities, one for each database target. Using these probabilities, a maximum-likelihood determination is made as to what the target actually is. It has been demonstrated that this technique can reliably determine the presence of up to three people, as well as distinguish between vehicle types (wheeled vs. tracked vehicles, for example).

A CNN-Based Approach to Target Recognition

Target recognition using a convolutional neural network (CNN) can outperform more traditional approaches. After training on synthetic images, it has proven useful for recognizing targets (such as battle tanks) in infrared images of real scenes. How realistic the synthetic images are matters greatly when trying to recognize real scenes from the test set because of the constraints of the training set.

Seven convolution layers, three max pooling layers, and a Softmax layer serve as the backbone of the overall CNN network architecture. Following the second, fourth, and fifth convolution layers are the max pooling layers. Prior to the final result, a global pooling average is applied. Leaky ReLU is used as the nonlinear activation function in all convolution layers.

{End Chapter 1}

Chapter 2: Computer vision

The study of how computers can derive high-level knowledge from digital pictures or videos is the focus of the multidisciplinary scientific area of computer vision. From a technological point of view, it investigates and attempts to automate activities that are within the capabilities of the human visual system.

Tasks associated with computer vision include techniques for obtaining, processing, analyzing, and comprehending digital pictures, as well as the extraction of high-dimensional data from the physical environment in order to create numeric or symbolic information, such as judgments.

Computer vision is a subfield of computer science that investigates the theoretical underpinnings of artificial systems designed to derive information from pictures. The visual data may be presented in a variety of formats, including video sequences, images obtained from several cameras, multi-dimensional data obtained from a 3D scanner or medical scanning equipment, and so on. The goal of the technical field known as computer vision is to implement the ideas and models it has developed in the process of building computer vision systems.

The fields of scene reconstruction, object detection, event detection, video tracking, object recognition, 3D pose estimation, learning, indexing, motion estimation, visual servoing, 3D scene modeling, and image restoration are all sub-domains of computer vision. Other sub-domains of computer vision include 3D scene modeling.

Computer vision is a multidisciplinary study that examines how computers can be programmed to extract high-level knowledge from digital pictures or movies. This area focuses on how computers can be taught to comprehend what is being shown to them. From the point of view of engineering, the goal is to find ways to automate operations that can already be done by the human visual system. Computer vision is a field of study in the field of information technology that focuses on applying existing theories and models to the process of building computer vision systems.

In the late 1960s, colleges that were on the cutting edge of artificial intelligence were the first to experiment with computer vision. Its purpose was to function in a manner similar to that of the human visual system, with the ultimate goal of imbuing robots with intelligent behavior. In the 1990s, several of the study areas that had been studied before became more active than the others. The study of projective three-dimensional reconstructions led to a deeper understanding of how to calibrate a camera. It became clear, with the introduction of optimization techniques for camera calibration, that a significant number of the concepts had previously been investigated by the discipline of photogrammetry's bundle adjustment theory. This came to light as a result of this development. This resulted in the development of techniques for doing sparse three-dimensional reconstructions of scenes using several photographs. Both the dense stereo correspondence issue and the development of further multi-view stereo approaches saw some degree of forward movement. Concurrently, many variants of graph cut were used in order to address the picture segmentation problem. This decade was especially significant since it was the first time statistical learning methods were used in practice to the task of recognizing faces in photographs (see Eigenface). The areas of computer graphics and computer vision have become more intertwined in recent years, which has led to a large rise in the amount of collaboration that has taken place between the two. This featured early forms of light-field rendering, panoramic picture stitching, image morphing, view interpolation, and image-based rendering. The area of computer vision has been given a new lease of life thanks to the development of algorithms based on deep learning. The accuracy of deep learning algorithms on numerous benchmark computer vision data sets for tasks ranging from classification to optical flow has exceeded that of earlier approaches. These tasks include segmentation of images and optical flow.

Solid-state Computer vision is strongly connected to a number of other disciplines, including physics. The vast majority of computer vision systems are based on image sensors, which are devices that are able to detect electromagnetic radiation. This radiation is commonly manifested as either visible or infrared light. Quantum physics was used in the development of the sensors. The scientific discipline of physics provides an explanation for the method through which light interacts with surfaces. The behavior of optics, which is a fundamental component of the majority of imaging systems, may be explained by physics. In order to offer a comprehensive knowledge of the process by which a picture is formed, sophisticated image sensors need the use of quantum mechanics. Computer vision may also be used to solve a variety of measurement issues that arise in physics, such as those involving the motion of fluids.

The scientific discipline of neurobiology, more especially the investigation of the biological visual system. The eyes, the neurons, and the brain regions that are dedicated to the processing of visual inputs in both humans and diverse animals have been the subject of a substantial amount of research over the course of the last century. As a consequence of this, a simple but intricate description of the way actual vision systems function in order to complete various vision-related tasks has emerged. As a consequence of these findings, a subfield of computer vision has emerged in which artificial systems are meant to