Portal de Periódicos da CAPES

Drone detection and radar‐cross‐section measurements by RAD‐DAR

2019; Institution of Engineering and Technology; Volume: 13; Issue: 9 Linguagem: Inglês

10.1049/iet-rsn.2018.5646

1751-8792

Álvaro Duque de Quevedo, Fernando Ibañez Urzaiz, Javier Gismero Menoyo, A. Asensio-López,

Target Tracking and Data Fusion in Sensor Networks

IET Radar, Sonar & NavigationVolume 13, Issue 9 p. 1437-1447 Special Issue: Selected Papers from the 2018 International Conference on Radar (Brisbane, Australia)Free Access Drone detection and radar-cross-section measurements by RAD-DAR Álvaro Duque de Quevedo, Corresponding Author Álvaro Duque de Quevedo aduque@gmr.ssr.upm.es Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this authorFernando Ibañez Urzaiz, Fernando Ibañez Urzaiz Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this authorJavier Gismero Menoyo, Javier Gismero Menoyo Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this authorAlberto Asensio López, Alberto Asensio López Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this author Álvaro Duque de Quevedo, Corresponding Author Álvaro Duque de Quevedo aduque@gmr.ssr.upm.es Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this authorFernando Ibañez Urzaiz, Fernando Ibañez Urzaiz Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this authorJavier Gismero Menoyo, Javier Gismero Menoyo Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this authorAlberto Asensio López, Alberto Asensio López Grupo de Microondas y Radar, Information Processing and Telecommunications Center, Universidad Politécnica de Madrid, Avenida Complutense 30, Madrid, SpainSearch for more papers by this author First published: 01 August 2019 https://doi.org/10.1049/iet-rsn.2018.5646Citations: 8AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract RAD-DAR is a frequency-modulated-continuous-wave radar demonstrator, working at X-band, completely designed and constructed by the authors' research group following the ubiquitous-radar concept. After introducing its main hardware and software blocks, with special emphasis to its off-line signal processing and data processing, this study presents the results of a drone detection test, where RAD-DAR achieved a DJI-Phantom 4 detection and tracking at a range up to 3 km. The system performance is discussed with different flights including an attack manoeuvre, and a free flight, which helps to highlight their advantages in surveillance tasks due to the RAD-DAR staring nature. Furthermore, range, speed and azimuth accuracies are discussed, considering the drone global positioning system data. Finally, this study shows a statistical radar cross-section study based on the processed data of the drone and attempts to classify this target by means of Swerling models. 1 Introduction Unmanned aerial vehicles (UAV) have ceased to be a defence-exclusive element, to become an everyday object in anyone's life. They are used for commercial purposes in package delivery, in surveillance tasks, Earth observation, crops control, film shooting and, of course, they serve as an entertainment object as well. It is quite common nowadays to read about drones involved in more particular tasks such as communication networks creation or disease control. The current drone boom and their optimal state of development are evident. Drones are affordable goods and easy to use for an ever-increasing public. Anyone can become an UAV pilot, managing a cutting-edge technology object with a spectacular performance, without the need of a great outlay or many hours of practise. This, in fact, is precisely what makes drones to be high in customer appeal, but also a potential security-and-privacy threat. It seems to be too easy to take pictures of protected areas or infringing personal privacy, to create an alert likely to force the closure of an airport or airspace or even to cause an air accident, personal injuries or to spread panic about a potential terrorist threat. Thus, this frenetic UAVs development itself has led to a parallel growth of detection and shooting-down systems. Undoubtedly, among other detection techniques employed in the last decade (e.g. audio-based, video-based [1], radio-frequency-based, thermal sensors based and wireless fidelity-based [2] systems), radar technology is one step ahead [3] due to its weather-regardless performance. However, micro-drone detection and their early warning are major challenges to radar systems [4] because these small targets are made of low-reflective plastic materials, and they are capable of flying at low speed and at ground level. This makes them very low radar-cross-section (RCS) targets, difficult to detect in high clutter environments. Conventional surface surveillance radars, with mechanical or electronic exploration, have severe constraints in terms of detection of these kinds of targets. These mentioned systems work with large bandwidth and high range resolution in order to minimise the superficial clutter impact, in which power is proportional to the resolution according to the expression [5, p. 2.17] (1) where is the surface clutter RCS; is the surface reflectivity; and is the illuminated area, which can be expressed as (2) where is the range resolution; R is the distance to the surface; is the azimuth beamwidth of the antenna and is the angle of incidence of the signal on the surface. In addition to the above, conventional surveillance radars are required to use relatively high data refreshment rates (0.5–1 Hz) because of the short time of the drones in the coverage area, during which an alarm must be triggered, and the drone guidance systems must be inhibited. This means that mechanical exploration radar is required to rotate up to 30–60 rpm. This high rotation speed of the radar antenna has two negative consequences for the system. On the one side, the scanning effect increases because the dwell time is low, thus the spectral width of the clutter increases as well. In this regard, it is noted that the clutter cut-off frequency can be expressed as [5, Ch. 2.4] (3) where is the dwell time during one antenna scan. On the other hand, due to the high rotation speed, the number of processed echoes during a dwell time is low. Consequently, the system capacity to deal with clutter by means of coherent processing is limited because of the low available echoes energy, and the system range is restricted as well. This paper introduces our radar with digital array reception technology (RAD-DAR): a small, quick-to-deploy and low-powered radar demonstrator system, based on the ubiquitous-radar concept [6]. Our first experimental results [7] showed the feasibility of achieving strong range–speed association with the system demonstrator, focusing on vehicle detection. After that, [8] presented the first RAD-DAR results on drone detection with very promising outcomes. This paper introduces our RAD-DAR system and its main hardware (Section 3) and software (Section 4) blocks, with special emphasis on the Doppler processing and digital beamforming implementation. After that, this paper discusses the theoretical capability of RAD-DAR to detect a micro-UAV in Section 5 and presents the new experimental results of a complete test on drone detection in Sections 6 and 7, enhancing the RAD-DAR successful detection and tracking of a micro-UAV flying at a range of 3 km, conducting an attack manoeuvre. The collected data are also employed to discuss the system performance by means of its range, speed and azimuth errors. Finally, the data collection is used in Section 8 to study the drone RCS behaviour. 2 Ubiquitous-radar concept The system presented in this paper, RAD-DAR, is a radar demonstrator designed and constructed following the ubiquitous concept guidelines [6]. These types of systems, also referred to as staring, holographic or persistent systems include a transmitting antenna which constantly illuminates the entire monitored area, and N receiving antennas, similar to the transmitting one. Each receiving antenna has its own reception channel with its corresponding analogue-to-digital converter. Thus, RAD-DAR is a surveillance radar system with digital array reception (Fig. 1). Fig. 1Open in figure viewerPowerPoint Ubiquitous-radar operating principle The system provides data cubes to the signal processor, in which three dimensions are: the number of range cells (Nr), the number of reception channels (Na) and the number of periods to be processed (Nd), as can be seen in Fig. 2. Fig. 2Open in figure viewerPowerPoint RAD-DAR data cube By means of this operation mode, there is no limitation in terms of dwell time, since there is no scanning. As a result, the radar range is theoretically unlimited when the system deals with thermal noise, and in practise the required range can be achieved with relatively low transmitted power. Conversely, there is a downside concerning the receiving antenna: its gain is low because its azimuth beamwidth is the same as the whole monitored area (100°). The data refreshment rate can be high as well, inasmuch as this operative magnitude does not depend on any scan speed, but rather only on the amount of energy that the system must integrate to achieve required range coverage. A key advantage of this system architecture lies in the absence of scanning, leading to a narrow clutter spectral bandwidth. There is not any scanning effect: after a coherent signal processing (Doppler processing), clutter signals are confined to a one or two Doppler bins. Furthermore, these Doppler bins represent very low speeds because the dwell time is up to a 100 times higher than the dwell time in a conventional surveillance radar system. Moving targets virtually compete only with thermal noise so it is not necessary to use high bandwidth because a high range resolution is not required. Indeed, a too high range resolution can actually be harmful, leading to target cell migration during the dwell time. This migration would result in signal-to-noise ratio (SNR) losses and, thus, loss of range coverage. Nevertheless, a good range resolution is desirable to obtain accurate position measurements, which is of crucial importance in the association process at the radar data processor. Finally, another advantage of this system has to be remarked: in contrast with conventional systems, which provide only two target dynamics measurements (range and azimuth), a ubiquitous radar with Doppler processing provides plots with information about range, azimuth and radial speed of the target. Furthermore, these speed measurements are more accurate than the ones provided by a typical traffic control radar system. This leads to a very precise plot association capability, even in complex environments, and the possibility of implementing tracking filters with the available speed data. 3 Introducing RAD-DAR: hardware description Our RAD-DAR, showed in Fig. 3, implements the ubiquitous-radar concept discussed in Section 2 employing an 8-channel digital array which receives the echoes, digitises and stores them for further off-line processing. The system generates multiple simultaneous reception beams, by means of a digital beamforming process which allows it to achieve the required azimuthal coverage. This is the reason why our radar system is nicknamed as RAD-DAR. Fig. 3Open in figure viewerPowerPoint RAD-DAR system Fig. 4 shows a diagram of the RAD-DAR demonstrator system. It uses a frequency-modulated-continuous wave (FM-CW, Fig. 5) on a frequency band centred at 8.75 GHz (X-band) with a maximum bandwidth of 500 MHz. This waveform enables the system to achieve high range resolution (below 1 m) using sampling rates well below the instantaneous bandwidth [9]. A programmable signal generator (Direct Digital Synthesiser (DDS) + X-band phase-locked loop) provides this waveform (with selectable parameters as its bandwidth and its repetition frequency), clock and trigger signals. A transmitter amplifies the signal up to 5 W and generates the local oscillator sample for subsequent demodulation. Fig. 4Open in figure viewerPowerPoint RAD-DAR hardware Fig. 5Open in figure viewerPowerPoint FM-CW waveform The 8 receiving antennas, arranged in an 8 × 8 elements array (170 × 180 mm2), and the transmitting antenna, 8 elements array (40 × 180 mm2), are designed with microstrip technology [7]. The transmitted beam has a beamwidth of 11° in elevation and 86° in azimuth. Fig. 6 shows these azimuth and elevation beams. The received synthesised beams, by means of digital beamforming, have a beamwidth of 11° in elevation and 20° in azimuth. Fig. 6Open in figure viewerPowerPoint RAD-DAR antenna radiation pattern The signal from eight receivers is acquired by a commercial 8-channel digitiser [10] connected by a PCIe port to a computer, where the off-line signal and data processors are implemented using MATLAB software. It is important to highlight the special care to preserve the acquisition synchronism in order to perform a further coherent processing. For that purpose, the digitiser is also connected to the RAD-DAR clock and trigger signals. The system also has a power supply, which provides the required voltage to the transmitter and the receiver. The entire system has a weight of <15 kg and can be deployed and set up by a single person in <15 min. 4 Doppler processing: software description A MATLAB script has been developed to carry out the digitiser control. This signal acquisition subsystem captures data arranging them into cubes for a posterior three-dimensional (3D) coherent processing. As can be seen in Fig. 2, after that processing, the three cube dimensions will lead us to obtain information about range, Doppler frequency (speed) and azimuth. Fig. 7 shows a block diagram with a quick description of the radar software stages: acquisition subsystem, radar signal processor and data processor. Fig. 7Open in figure viewerPowerPoint RAD-DAR software The off-line radar signal processor, which is implemented with MATLAB, works with the data cubes (Fig. 2) generated by the acquisition subsystem. This signal processor performs a 3D processing [2D fast Fourier transform (FFT), beamforming and monopulse] and a detection-and-plot-extraction stage. After that, the off-line radar data processor groups these plots into tracks of the different targets in the captured scene. 4.1 2D fast Fourier transform Owing to the FM-CW waveform employed, targets range information lies in the 'beating frequency', which is the difference between transmitted and reflected frequencies, and can be extracted by means of a classic homodyne receiver scheme [7]. There is a linear relationship between target range R and beating frequency fb given by (4) where fm is the modulating signal frequency and Δf is the modulating bandwidth. The radar processor applies an FFT to the Nr dimension of data cubes (see Fig. 2), obtaining the beating frequency, which can be converted to range information by means of (4). A second FFT, performed over the Nd dimension, provides Doppler data, which are easily converted to speed information. 4.2 Beamforming The signal processor receives eight signals coming from the eight channels and synthesises five reception beams corresponding to pointing angles θ between −40° and +40°, in order to achieve the required azimuthal coverage. A phase increment is applied to each received signal, according to the inter-element spacing and the pointing angle [7]. The phase-shift compensation for each channel is (5) where θ is the steering angle, n is the channel and Δx is the separation between receiving antennas in the azimuthal plane. The system generates a range–Doppler matrix for each pointing angle by adding the eight phase-shifted signals. Fig. 8 shows the five receiving synthesised beams, obtained from real measurements of the receiving antenna in an anechoic chamber [7]. A Taylor window has been applied in order to reduce the secondary sidelobes level. The beams gain is normalised in this representation. Fig. 8Open in figure viewerPowerPoint RAD-DAR synthesised beams by beamforming 4.3 Monopulse To measure the targets azimuth and reject detections which do not correspond to a processed beam, the signal processor performs a monopulse technique. A difference beam has been designed for each channel. After generating the sum signal at the beamforming process, a difference signal is obtained for each pointing angle θ by subtracting channels 5–8 signals from channels 1–4 signals. The quotient between the magnitudes of difference signal (Δ) and sum signal (Σ) can be used to obtain a monopulse function. A number of mathematical expressions can be applied in order to obtain a valid monopulse function [5, Ch. 9.2], which can be used by the system to refine the azimuth information of each detection. The most important requirement regarding this expression is that a monopulse function must be an injective mathematical function, in order to be invertible. In this manner, the system can find the azimuth value corresponding to a single monopulse function value. A valid expression for a monopulse function e is (6) where δ is the phase difference between both sum and difference signals (or the phase of the quotient). Fig. 9 shows both sum and difference patterns for a pointing angle θ = 0°, and the corresponding monopulse function, obtained with expression (6), with emphasis on the region of interest inside the synthesised beamwidth. It can be seen how the difference beam covers the sidelobes of the sum beam, in which targets are detected. Fig. 9Open in figure viewerPowerPoint Sum and difference signals. Monopulse function Fig. 10 shows the monopulse function obtained with (6) for each receiving synthesised beams. These functions are used to translate a computed monopulse value of a plot, into an azimuth angle and to delete detections out of the valid monopulse values, that is, out of each beam coverage area. The lower slope of the monopulse functions at the ends of Fig. 10 (beams at −40° and 40°) will result in a loss of azimuth measurement accuracy at these extreme beams. Fig. 10Open in figure viewerPowerPoint Monopulse function 4.4 Decision and plot extraction The detection stage of the radar signal processor implements cell-averaging constant false alarm rate (CA-CFAR) technique [11] to obtain a binary cube with detections, referred to as plots. This stage compares the signal power at each range–Doppler cell with an adaptive threshold. For one cell, the threshold is calculated as the average signal power of m adjacent cells (reference cells), skipping n guard cells, which are fixed in order to avoid errors in the power estimation, caused by echoes from the target itself. After power CA, the decision threshold of one cell is calculated by multiplying the computed average power by a scale factor T a single value for the whole matrix, which is given by (7) where Pfa is the fixed false alarm probability and m is the number of reference cells [12, p. 599]. This CA-CFAR technique is implemented over both range dimension and Doppler dimension. The processor computes two thresholds: TD (Doppler dimension) and TR (range dimension) and employs the minimum of them to carry out the decision about a detection in a range–Doppler cell. In this way, the system manages to adapt in order to detect both targets extended in the Doppler dimension (e.g. humans) and targets extended in the range dimension (e.g. cars). Fig. 11 shows an example of 2D CA-CFAR window, with m = 8 reference cells (range and Doppler) and n = 1 guard cell (range and Doppler as well). The CA-CFAR window length, mD, mR, nD and nR can be selected at the signal processor according to the kinds of targets of interest in the monitored area. Fig. 11Open in figure viewerPowerPoint 2D CA-CFAR Each detection, usually composed of more than one cell, is processed and turned to a single-point detection (plot) by computing its centroid [13]. After this process, a set of plots are obtained, with data about range, speed, azimuth, monopulse value, received signal power and time. Fig. 12 shows an example of input to the decision-and-plot extraction stage: a range–Doppler matrix with raw data, obtained after 2D FFT and beamforming with pointing angle θ = 0°. Clutter can be appreciated at zero Doppler. Fig. 12Open in figure viewerPowerPoint Range–Doppler matrix. Raw data Fig. 13 shows the details of the range–Doppler cells corresponding to the monitored drone. As can be seen in that figure, in this example, the drone was detected by the system at a distance of 0.83 km, approaching with a speed of 36 km/h. Furthermore, Fig. 13 shows the result of the CA-CFAR-and-plot-extraction stages: the black contour represents the bins where there is a detection and the red point is the centroid of a set of detections, a plot. Fig. 13Open in figure viewerPowerPoint Range–Doppler matrix. Zoom to the drone 4.5 Data processor As can be seen in Fig. 7, the last stage of the RAD-DAR software is the radar data processor. A simple tracking algorithm is applied to the complete set of plots corresponding to a collection of processed cubes, in order to associate the detections in tracks that belong to the targets in the captured scene. The implemented tracking algorithm consists of associating plots to track according to the distance from each real plot to a predicted plot, based on the history of each track, at any one point of time. The algorithm considers the plots azimuth, range, radial speed and time data. The predicted plot estimation is simply based on two assumptions: target speed and azimuth are constants. Therefore, range variation is linear with the speed and predicted using the measured speed of a previous plot and the time interval from one cube to another. A new real plot can be associated to a track when it is inside a 3D window (range, speed and azimuth) around the prediction. This window size can be selected in the RAD-DAR data processor, but all the tracks showed in this paper have been obtained using 2 m × 2 km/h × 10°-sized windows. When a single plot is close enough to one track prediction, it is directly associated, but when there are multiple possible associations, the data processor applies the Munkres' variant of the Hungarian association algorithm [14], according to which the assignment solution is the one that minimises the total cost or distance for the problem. The distance is given by [15, p. 9] (8) where R, V and θ are the values of range, radial speed and azimuth, respectively, obtained from the real plot (candidate for the association); PR, PV and Pθ are the predictions of range, radial speed and azimuth of the considered track; and σR, σv and σθ are estimations of measurement errors committed by the system. The great range–speed association capacity of the RAD-DAR allows us to simplify the association algorithm as much as it has been exposed above, with very good tracking results. No tracking filters are currently implemented in the data processor. Both the high data refreshment rate and high range and Doppler measurement accuracies allow the system to associate plots without any filtering, with a false association probability practically zero. By means of this data processor, all the captured scenes can be cleaned, and false alarms deleted. That allows the system to work with very high false alarm probability at the signal processor, up to 1×10−3, without degrading the system capability to isolate one track in order to study only the plots of the target of interest, as will be seen in next sections. As an example of the signal processor and data processor outcomes, Fig. 14 shows a range–Doppler matrix with raw data (after 2D FFT and beamforming) corresponding to one cube, number 168, of a 302-cube scene (about 151 s), where a drone is describing a circular trajectory. The signal processor has worked with five synthesised beams centred at staring angles of −40°, −20°, 0°, 20° and 40°, as shown in Fig. 8, and the false alarm probability has been set to 1×10−3 at the CA-CFAR stage. Fig. 14Open in figure viewerPowerPoint RAD-DAR data processor in operation: raw video Fig. 15 shows the same range–Doppler matrix after the detection-and-plot-extraction stages. A significant number of false alarms (soft-red dots) are showed, due to the high false alarm probability. The tracking algorithm is applied to the collection of cubes, with the 2 m × 2 km/h × 10°-sized window previously described. The red-dotted line shows the extracted track of the drone, compared with the global positioning system (GPS) data, in blue. This figure, and all the results in Section 7, shows the tracks of extraction capability of the system, allowing us to extract the plots of the drone on every path, reducing the false alarm probability from 10−3 or 10−2, before the tracker, to zero (no false alarms) after the tracker. Fig. 15Open in figure viewerPowerPoint RAD-DAR data processor in operation: detections 5 Drone detection test setup 5.1 Target characteristics A DJI-Phantom 4 [16] was employed to perform the drone detection experiment described in this paper. It is a commercial micro-UAV designed for both private and professional use. Table 1 summarises its main characteristics. Table 1. DJI-Phantom 4 main specifications weight (propellers and battery included) 1380 g diagonal length (propellers not included) 350 mm maximum speed 72 km/h maximum flying height (above sea level) 6000 m positioning system GPS/Glonass maximum flying time (battery life) 28 min maximum range (remote control range) 5 km The main material of the drone body, as well as the propellers, is plastic, and the four engines are made of aluminium. It has also a camera at the bottom of the main structure, made by aluminium covered by plastic. A considerable amount of the literature has been published on drone cross-section. Farlik et al. [17], Schroder et al. [18] and Li and Ling [19] show results on RCS obtained by means of anechoic chamber measurements or simulation, of a DJI-Phantom 2, which is similar to Phantom 4 in terms of weight, size, materials and shape. There is a view widely held in these papers that a Phantom drone may be modelled as a Swerling 1 target (SW1) with an average RCS between 0.01 (−20) and 0.35 m2 (−4.6 dBsm), depending on the propellers movement, the transmitted wave polarisation and frequency, the elevation angle and other factors. DJI-Phantom 4 allows us to record its telemetry data (e.g. GPS coordinates, range and speed) during a flight, with a 10 Hz sample frequency. These data, which are easy to export for further processing, are helpful in order to carry out a quick extraction of the drone tracks from the radar data processor results, at the first experiment stages. Similarly, the drone GPS data are employed to estimate measurement errors, by comparing the RAD-DAR tracks with the on-board records. 5.2 Field-test scenario The chosen location for the experiment was a ranch situated at a village in the province of Ávila, Spain, with geographic coordinates 40°49′47.3″N 4°48′00.4″W [20]. The system was deployed over an emplacement with entirely free line of sight extended up to 5 km. In our tests, the drone described different trajectories including free flight (driven by a pilot, by means of the manual mode of the drone remote control), attack manoeuvre (linear trajectory against the radar system), circular path (the drone describes an orbit around a selected point) and radial trajectories (round trips with different azimuth angles). The main analysis in this paper deals with the attack manoeuvre, in which the drone described a return flight from a range of 3.1 km (Fig. 16), in automatic 'Go to Home' mode. This return linear flight is the best to assess the radar performance (range, speed and azimuth errors), since drone managed to keep a quasi-constant trajectory with low-speed deviation by means of that automatic control mode, as can be seen in Table 2. Furthermore, with this flight, the RAD-DAR capability to detect and track a small target, up to 3 km away, with only 5 W transmitted power will be also revealed. Table 2. Field tests: drone attack manoeuvre GPS data Flight control mode Auto (Go to Home) flight time 320 s average speed −34.65 km/h STD of the speed 0.26 km/h average altitude (above ground level) 30.84 m STD of the altitude 0.09 m Fig. 16Open in figure viewerPowerPoint Field-test map Fig. 17 shows the RAD-DAR system prepared for the test. This selected scenario is usually populated by some birds of prey, which will be