CC BY
26
METHOD OF AUTOMATIC PEDESTRIAN RECOGNITION IN ROAD SCENE BY MICRO-DOPPLER SIGNAL FOR SELF-DRIVING VEHICLE RADAR SYSTEMS
DOI 10.24411/2072-8735-2018-10302
Andrey V. Pluchevskiy,
JSC "Cognitive", Moscow, Russia; Tomsk State University of Control Systems and Radioelectronics (TUSUR), Tomsk, Russia, pluch.andry@gmail.com
Keywords: micro-Doppler, Fourier transform, automatic target recognition, pedestrian recognition, cadence diagram, threshold, automotive radar.
This paper represents the method of automatic pedestrian recognition by a high-resolution Doppler spectrogram unique characteristic. The recognition is performed in a road scene with moving cars on the background. The Doppler spectrogram is regarded as a two-dimensional radar image. Taking into consideration the features of a pedestrian micro-Doppler signal, the processing of a two-dimensional radar image is reduced to a one-dimensional threshold. The algorithm has been developed for operating in a sliding window mode alongside the continuous acquisition of data on the target Doppler spectrum. The proposed technique has been developed analytically and does not require the use of machine learning and deep learning algorithms. Experimental research was conducted on two types of road scene objects: pedestrians and automobiles. The experiment results showed that the proposed method can distinguish a pedestrian from a moving automobile even if a Doppler bandwidth is similar. The work investigates the detection probability of the proposed method according to the signal-to-noise ratio and the false alarm probability. It enables to set requirements for a radar system on a design stage or to evaluate the possibility of applying the method in existing systems. The method is suitable for application in the radar and computer vision fields. The proposed technique was developed for use in driver assistance systems and the automotive vehicle industry to recognize pedestrians and take necessary measures for collision avoidance.
Information about author:
Pluchevskiy Andrey Vladimirovich, Junior development engineer, radiolocation department, JSC "Cognitive", Moscow, Russia graduate student, assistant, Tomsk State University of Control Systems and Radioelectronics (TUSUR), Tomsk, Russia
Для цитирования:
Плучевский А.В. Метод автоматического распознавания пешеходов в дорожной сцене по сигналу микро-Доплера для радиолокационных систем беспилотного автотранспорта // T-Comm: Телекоммуникации и транспорт. 2019. Том 13. №8. С. 51-59.
For citation:
Pluchevskiy A.V. (2019). Method of automatic pedestrian recognition in road scene by micro-Doppler signal for self-driving vehicle radar systems. T-Comm, vol. 13, no.8, pр. 51-59.
Introduction
In recent years, there has been a steady trend of radar application in the development of self-driving car systems [1,2]. Radars arc generally applied for target range and velocity estimation. Also, radars are used for surrounding area mapping as a part of a complex sensor system, providing different multi-dimensional information about a road scene [3,4]. Herewith, digital methods are used both in the radar systems and computer vision systems for signal processing and automatic target recognition (ATR). Detection of pedestrians against other road scene objects on the background is one of the important problems for a collision-avoidance system.
According to [5]. the development of the radar target recognition method consists of three stages;
1. The selection of objects and scenarios for recognition;
2. The selection of a feature extraction algorithm and a measurement technique;
3. The selection of the decision-making criterion.
This paper continues the previous work [6] and concludes the consideration of all three stages.
The European New Car Assessment Programme (EURO NCAP) considers four main classes of road scene objects; a vehicle, a pedestrian, a cyclist, and a motorcyclist. This research is focused on pedestrian recognition when observing a road scene with moving vehicles and pedestrians, A cyclist and a motorcyclist are not studied to simplify the analysis of target features. Nevertheless, the proposed technique is supposed to be a part of a complex recognition system.
Methods based on Doppler signatures [7,8] and RCS signatures [9] are commonly used to recognize a pedestrian against the background of other road scene objects by one frame. These techniques provide recognition results in a short time (milliseconds, dozens of milliseconds). Besides, there are techniques that extract additional information from the signal slowly varying characteristics obtained by high-resolution measurements of the micro-Doppler (m-D) effect [10-22]. Radar sensor application involves the use of known radar signal processing methods and the established terminology. Thus, for radar target signal detection in a background interference, the following basic concepts are used; the probability of detection (PD), the probability of false alarm (PF), and signal-to-noise ratio (SNR). The performance of signal detection is traditionally represented as PD dependence on the SNR and PF [23].
The recognition task is solved by machine learning and deep learning approaches in the significant list of publications [11-22]. The performance of these algorithms is determined by accuracy, correct classification rate, and confusion matrix obtained in experiments.
It should be noted that some algorithms have quite a high accuracy of about 99% [12j. However, there has been insufficient research into the accuracy dependence on the SNR and Pj. and requires additional studies.
The application of automotive radar sensors for unmanned vehicles implies restrictions on the cost and the hardware computational load [1,2]; also, there are strict requirements for the transparency of decision-making algorithms [7] and unambiguity oftarget recognition regardless of a road scene conditions.
Therefore, this paper represents the pedestrian recognition method wherein the processing of a Doppler spectrogram as a two-dimensional (2D) radar image is consequently reduced to the
one-dimensional (ID) threshold, calculated according to the Neyman-Pearson criterion, which characteristics are well-studied.
1. Genera] description of the recognition task
Fig. 1 represents an example of a dangerous road scenario, showing the necessity of recognition. The pedestrian has to bypass some irregular obstacle (e.g., a puddle, repair works, a dog, etc.) on a sidewalk. The unmanned vehicle must recognize the moving target as a «pedestrian» and considering their ability to move unpredictably avoid possible collision and ensure safety.
t
'A
ft
Fig. 1, The road scenario where a pedestrian has to step on a roadway to avoid an obstacle
In mathematical terms, the recognition task is defined as follows. Let b={bi, b2...b„} be a plurality of objects. Suppose A = {Ai, Aj, ...A,,,} is a predefined class alphabet that object b belongs to and x={jo, xi ...Xk} is a dictionary of features that describes objects of plurality b. According to these terms, recognition is function F that matches given object by its features .v to class A
4 =F(b =
(0
For the developed method, objects belonging to plurality b denote radar targets. The predefined classes are «pedestrian» and «automobile». Features x are a set of spectral components extracted from a micro-Doppler signal, which allows distinguishing a pedestrian from an automobile in a road scene. The essential features are 2D spectrum components that represent the periodicity of a radar image over the time axis (Fig, 2),
a) 3 Pedestrian m-D 2D Spectrum ^, Car m-D 2D Spectru m
j H 44
2 S 25
42 ça
1 i I15 5 2 L5 □
40 |
s |
■s »
o >> i u OS 36 £
0.5 ■
m 34
0 __ 9 ■jt 0 ■ 32
Cycles per s
Cvdes per s
Fig, 2. (a) Spectrum of a pedestrian m-D Signal; (b) Spectrum of an automobile m-D signal
A Doppler spectrogram is regarded as a 2D radar image and considered as a miero-Doppler signal. To obtain a 2D spectrum, a 2D fast Fourier transform (FFT) of a Doppler spectrogram is used:
ff-l M 1 _/2jr(iiU—)
= »"• C2)
where k,l are spatial frequencies in n and m directions, respectively; - a 2D spectrum of a 2D radar image i[n,/w]; N and M are the number of samples in n and m directions, respectively.
The horizontal axis of 2D spectrum S[£./] (Fig. 2a) shows the periodicity of a micro-Doppler signal, and its physical meaning is a frequency of arm and leg swing.
2. Recognition algorithm
Let us reduce a recognition process to a signal detection process including a background interference, where the interference is a response from a vehicle and noise. Further, the probability of detection Pd will define the correctness of mapping a given radar target bj to class A, — «pedestrian» (1), if the target is a pedestrian. In other words, the Pn is considered as a recognition probability in terms of recognition. Let us apply the principles of the optimum signal delecting from [24]. A quadrature receiver should be used to detect a stochastic signal with a random phase in the background noise, and to detect a signal with the presence of interfering signals, a filter must be used.
Since the essential feature is a spectral component that lies only on one axis of the 2D spectrum, it is sufficient to perform a ID Fourier transform for all the Doppler frequency slices and then to average the result over each slice of the cadence frequency.
pFp C[kjx] ■-
________
_-> _
-*
Fig. 3. Step I: Fourier transform for all the Doppler frequency slices, where n is time, m - Doppler frequency, k - cadence frequency
For each wz-th Doppler Irequeney slice oi spectrogram a[>i.m\. the mean value of the slice is subtracted, then FFT in time direction ti is executed. An FFT from the spectrum is called a cepstral analysis and result image C[k,m] is called a cadence diagram. Such processing is provided in [15], succeeded by applying machine learning algorithms for recognition.
Step 2.
Considering that the periodicity in each Doppler frequency slice would remain virtually unchanged for the constant pedestrian speed [6], cadence diagram C[k,m] can be averaged over each slice of a cadence frequency k. The result is a ID cadence frequency vector.
eft»]
Uile
Fig. 4. Step 2: The averaging of cadence frequency slices, where k is a cadence frequency and m is a Doppler frequency
Step 3.
The threshold can be applied to the resulting ID vector. The threshold is defined by the Neyman-Pearson criterion. If a Doppler spectrogram belongs to a pedestrian, then within a 1D spectrum, a response is observed on the frequencies from 1 to 2.5 Hz.
The detection threshold value is defined by PF [25] and it is calculated as follows:
h - yj-2LnPr , (3)
cadence frequency
Fig. 5. The thresholding of a 1D cadence frequency vector, where h is a defined threshold level
Figure 6 represents the flowchart of the algorithm, uniting all the steps.
Fig. 6. The recognition algorithm flowchart
Step I.
The mean value of arm swing frequency is 0.5 Hz, while a pedestrian moves at 1 km/h speed, according to [26]. The arm swing frequency is 1.25 Hz for 7 km/h speed. The swinging of both arms is observed in the spectrogram without distinguishing the left hand from the right one. This means that a cadence frequency is doubled; 1 and 2.5 Hz for 1 and 7 km/h, respectively. Consequently, the observation time should be no less than Is.
Figure 7 illustrates the situation showing bow under the conditions of oncoming traffic the observed velocity of a grey vehicle decreases and becomes zero when passing a vertical radar antenna plate. In other words, the radial acceleration is observed. The velocity change of the front, middle, and back parts of a car happens non-uniformly because a car is a distributed target. This is sIiowti in Fig. 8.
3. Conditions of the algorithm
Since the recognition algorithm is based on the changing of Doppler frequency periodicity, the radar velocity resolution is the most important parameter that determines the possibility of applying the method.
Sv=-
TF%fc
(4)
where c is the speed of light, Tp is the period of the frame, and /, is the carrier frequency.
The velocity resolution should be such that it was possible to distinguish limb movements from the body movement. The peak velocity of anus and legs movements is approximately equal to the double body velocity. This can be seen in velocity profiles in [27-281. Thus, the velocity resolution should be no lower than the minimum defined pedestrian velocity (e.g.: if vPmin=\ km/h, then Sv<l km/h).
The power level of the signal reflected from limbs is about 15-20 dB lower than of the one from a body; hence, the resolution should be increased by 2 times.
The radial component of a target velocity is measured in order to obtain a Doppler spectrogram. The radial velocity is defined as Vr — VCOS0. An angle 0 is the angle between a target
velocity vector and a radar radius vector. In case a target moves in parallel to a radar sightline, 0 is also a radar target direction of arrival (DOA), which is shown in Fig, 7, If 0-60°, then vr='A v and taking into account the above-mentioned, the velocity resolution should be additionally increased by 2 times.
Since the observation time is about 1 -2s, overtaking or oncoming traffic leads to rapid DOA changing and causes a significant change of a Doppler spectrogram form.
time, s
7
Fig. 8. The Doppler spectrogram of the automobile thai crosses the radar antenna plate. The red lines show the car front part DOA 0 at different moments
This spectrogram (Fig. 8) has periodicity over the time axis, and it leads to errors on the decision-making step because the cadence diagram has the same spectral components as a pedestrian one. Also, acceleration influence on a Doppler spectrogram form is discussed in [6].
Based on the foregoing, the following conditions have to be met in order to obtain a correct recognition result;
a) The velocity resolution of a radar should be 4 times higher than the defined minimum pedestrian velocity: Sv < 'A vmin,
b) The operating range of the target azimuth should be limited by -60°<6<60°, where the characteristics of the cosine are close to the linear.
4. Analytical characteristics and simulation result
The detection probability of a stochastic signal with a random phase in the background noise [25] is calculated according to the following analytical expression:
PD = Q(yj2SNR,h), (5)
where h is the detection threshold, SNR is the signal-to-noise ratio, and Q(a,b) is a Marcum Q-function [25] defined by
Q(a,b)= £
r exp
h(ar)dr
Fig. 7. A typical road scene where automobiles move toward each other
where Iq(x) is a modified Bessel function.
The curves calculated by (5) are shown in Fig. 11 with the solid lines.
The model used for the simulation is a Doppler spectrogram calculated by the velocity profiles from [27]. An additive white Gaussian noise (AWGN) was added to a spectrogram to provide a defined the SNR. A window with the length of 1.5s slides along the spectrogram and sets the input data for the recognition algorithm (Fig, 6). For the given Pr, the threshold is computed, and the decision on whether the target is a pedestrian or not is made for each step of the slide window.
The result of the simulation is represented in Fig. 11. The simulated curves do not fully match analytical curves calculated by (5). The curves obtained by the simulation follow the shape of analytical ones but lay 2,3 dB to the right. This shift can be interpreted as a decrease at the constant SNR. Or, alternatively, the proposed algorithm requires an increase of the SNR to maintain the same Pa value. This decrease is explained by the fact that an FFT uses a sine wave and cosine wave forms and matches them. Since a slice of a Doppler spectrogram is not a harmonic signal and closer to a pulse signal with sin(jr)/jr form, there is inconsistency with the analytical calculation conditions, and hence, the P» decreases. Despite the discrepancy, the proposed method performance can be calculated by (5), taking into account the SNR decrease (-2.3dB).
5. Experimental research
In the experiment, the radar evolution board A WR1243 Boost of Texas Instruments was used. The technical description is provided in [29]. The LFMCW fast-ramp technique was used for measurement with the following parameters: 79 GIIz carrier frequency, 4 GHz sweep bandwidth, 5 MHz sample rate of complex point ADC, 256 complex samples per chirp, 40 ¡.is chirp duration Tp, 156 gs chirp repetition interval, 256 chirps per one frame, 40ms total frame duration 7>, and 25 frames per second.
Signal pre-processing
To obtain a Doppler spectrogram, the algorithm represented in Fig, 12 was used. Beat signals obtained by each sweep are stored in the matrix for the whole frame. After the frame ends, the mean value of each fast time slice is subtracted from the slice to remove static clutter. Then FFT for each fast lime slice is computed to estimate a Doppler frequency. The averaging of each Doppler frequency slice is applied to the obtained matrix. The result is an instantaneous Doppler spectrum, which is used
Scenarios
Two types of road scene objects were studied in the experiment: a pedestrian (Fig. 13) and an automobile (Fig. 14). The pedestrian moves uniformly straight along the radar sight line with 3 different velocities: ~3, 4.5, and 6 km/li. Two types of a gait were studied: a normal type with natural arm swing and the one without swinging, when arms are pulled close to the body. The pedestrian started to move while being out of the radar field of view. Some of the representative experimental data of the pedestrian are shown in Fig. 13.
Fig. 13. On the left - the experimental scenarios for the pedestrian and the corresponding spectrograms are on the right side of the picture
Even without ami swing, the pedestrian Doppler spectrogram has a periodical structure formed by leg movements as it is shown in Fig. 13b. This allows using the proposed method when a human carries bags without swinging arms.
The automobile also moved uniformly straight along the radar sight line with 3 different velocities: =5, 10, and 20 km/h. The automobile started to move in front of the radar and then comes out of the radar view field as shown in Fig. 14.
MM
BW ,vftfffttf-'"j|
» ¿i ÏÎ III aw« tic, i i
Hg1""!' ■
» it M tl M M lí oie. s
o
model Pp=10
• exp. P[:=!0"6
model P=10~: F
• exp. PF-10'2
-6
r r i /
10 15 ZD 25
Fig. 14. On the left - the experimental scenarios for the automobile and the corresponding spectrograms are on the right side of the picture
Experiment results
The data obtained during the experiment is a set of Doppler spectrograms. The field of interest has been cut out of the whole spectrogram manually. The proposed recognition method was applied to each spectrogram as shown in Fig. 6.
SNR, dB
Fig. 15. The performance of the proposed recognition method: the dashed lines depict PD obtained by the simulation; the asterisks show Pa obtained experimentally. Red colour stands for P,=1 IT"; blue colour stands for /^=10""
The SNR for the experimental data was calculated as the mean value of all target signal points within a single spectrogram.
As a result of the experiment, the performance chart (Fig. 15) was obtained. The asterisks show the characteristics obtained by experimental data processing. The experimental data have been obtained not for a wide range of the SNR. Nevertheless, the experimental points lie on the simulated curves. This makes it possible to suppose that the detection performance of the proposed method is determined by (5) with an equivalent reduction of the SNR by 2.3 dB.
Conclusion
This paper provides the method of automatic pedestrian recognition in the road scene by a micro-Doppler signal for self-driving vehicle radar systems. An automobile and a pedestrian are treated as objects of the road scene. An input for the recognition technique is a segment of a Doppler spectrogram with duration not less than Is and with all target micro-Doppler signal bandwidth. The probability of recognition PD is 0.9 with /V=10~' and SNR=>20. The PD is estimated according to the SNR and Pp. The method considers features of a moving target Doppler spectrum; namely, a Doppler spectrum bandwidth of a moving automobile is comparable with a Doppler spectrum bandwidth of a pedestrian. Also, the conditions of the method are discussed. The velocity resolution of the radar must be not lower than V* of the defined minimum pedestrian velocity. The target DOA must be limited by ±60°.
The signal processing is reduced to the 1D detection threshold defined by the Neyman-Pcarson criterion. Machine learning and deep learning algorithms are not used within the method. The proposed method is considered to be used jointly with other recognition techniques, which provide a quick recognition result by one frame (~20ms). This method can be improved by using CFAR algorithms which provides an estimation of the threshold in the current conditions. The developed method can be applied in self-driving vehicle radar systems and driver assistance systems. The method advances the micro-Doppler signal analysis, improves the characteristics of the existing pedestrian recognition techniques, extends an application field of radar recognition systems and establishes the role of radars among other sensors in complex computer vision systems.
The future research supposes the development of a complex automatic recognition method for all four road scene object classes: a vehicle, a pedestrian, a cyclist, and a motorcyclist.
Acknowledgement
The author thanks C-Pilot team, Manokhin Gleb, Vciikanova Elena, Kostarev Aieksey for their contributions and especial thanks to Popov Yuriy for this research organizing.
References
1. Patole, S. M., Torlak, M„ Wang. D„ & Ali, M. (2017). Automotive radars: A review of signal processing techniques. IEEE Signal Processing Magazine, vol. 34, no. 2, pp. 22-35.
2. Hasch, J. (2015). Driving towards 2020: Automotive radar technology trends. The 2015 IEEE MTT-S International Conference on Microwaves for Intelligent Mobility ÍICM/M), pp. 1-4.
3. Kim, J„ Han, D. S„ & Senouci, B. (2018), Radar and vision sensor fusion for object detection in autonomous vehicle surroundings. The 2018 Tenth International Conference on Ubiquitous and Future Networks (ICUFN), pp. 76-78.
4. Steinbacck, J„ Steger, C., Holweg, G„ & Druml, N. (2017). Next generation radar sensors in automotive sensor fusion systems. The 2017 Sensor Data Fusion: Trends, Solutions. Applications (SDF), pp. 1 -6.
5. Gorelik A.L., Barabash Ju.L, Krivosheev O.V. (1990). Selekcija i raspoznavanie na osnove lokacionnoj informacii | Radar information based recognition and selection]. Moscow: Radio i svjaz, 240p. (in Russian)
6. Pluchevskiy A,V. (2019). Micro-Doppler feature extraction of pedestrian and automobile using two-dimensional fast Fourier transform. T-Comm, vol. 13, no,5, pp. 61-68.
7. Bartsch, A., Fitzck, F„ & Rasshofer, R. H. (2012). Pedestrian recognition using automotive radar sensors. Advances in Radio Science, vol. 10, no. 2, pp. 45-55.
8. Nanzer, J. A., & Rogers, R, L. (2009). Bayesian classification of humans and vehicles using micro-Doppler signals from a scanning-beam radar. IEEE Microwave and Wireless Components Letters, vol. 19, no. 5, pp. 338-340.
9. Lee, S., Yoon, Y. J„ Lee, J. E„ & Kim, S. C. (2017). Human-vehicle classification using feature-based SVM in 77-GHz automotive FMCW radar, IET Radar. Sonar & Navigation, vol. 11, no, 10, pp. 1589-1596.
10. Mcsloub, A., Abcd-Meraim, K., & Bclouchrani, A, (2017). Ground moving target classification based on micro-Doppler signature using novel spectral information features. The 2017 European Radar Conference (EURAD), pp. 255-258.
11. Zenaldin, M., & Narayanan, R. M. (2016). Radar micro-Doppler based human activity classification for indoor and outdoor environments, Proc. SP1E, [online] vol. 9829, no. 98291. Available at: http://dx.doi.org/IOJ 117/12.2228397 (Accessed 18 Feb.2019}
12. Cao, P., Xia, W„ & Li, Y. (2018). Classification of Ground Targets Based on Radar Micro-Doppler Signatures Using Deep Learning and Conventional Supervised Learning Methods. Radioengineering, vol. 27, no, 3,pp. 835-845.
13. Erol, B„ Amin, M. G., & Gurbuz, S. 2. (2018), Automatic data-driven frequency-warped cepstral feature design for micro-Doppler classification. IEEE Transactions on Aerospace and Electronic Systems, vol. 54,110.4. pp. 1724-1738.
14. Tivive, F. 11. C„ Phung, S. L., & Bouzcrdoum. A. (2015). Classification of micro-Doppler signatures of human motions using log-Gabor filters. IET Radar, Sonar & Navigation, vol. 9, no. 9, pp. 1188-1195.
15. Bjorklund, S., Petersson, H., & Hendeby, G. (2015). Features for micro-Doppler based activity classification, IET radar, sonar & navigation, vol. 9, no. 9, pp. 1181-1187.
16. Kim, Y., & Moon, T. (2015). Human detection and activity classification based on micro-Doppler signatures using deep convolutional neural networks. IEEE geoscience and remote sensing letters, vol. 13, no. I, pp. 8-12.
17. Erol, B., & Giirbuz, S. Z. (2016). Hyperbolically-warpcd cepstral coefficients for improved micro-Doppler classification. The 2016 IEEE Radar Conference (RadarConj), pp. 1-6.
18. Bjorklund, S„ & Rydell, J. (2017). Micro-Doppler classification with boosting in perimeter protection. Belfast, UK. The International Conference on Radar Systems, pp. 23-26,
19. Giirbuz, S. Z., Erol. B., ¿agliyan, B„ & Tekeli, B. (2015). Operational assessment and adaptive selection of micro-Doppler features. IET Radar. Sonar & Navigation, vol. 9, no. 9, pp. 1196-1204.
20. Du. R„ Fan, Y„ & Wang, J. (2018). Pedestrian and Bicyclist Identification Through Micro Doppler Signature With Different Approaching Aspect Angles. IEEE Sensors Journal, vol. 18, no. 9, pp. 3827-3835.
21. Li, W., Xiong, B., & Kuang, G. (2017). Target classification and recognition based on micro-Doppler radar signatures. The 2017 Progress in Electromagnetics Research Symposium-Fall (PIERS-FALL), pp. 1679-1684.
22. Fan, Y„ Du, R„ & Wang, J. (2018). Identification of Pedestrian and Bicyclist through Range Micro Doppler Signatures. IEICE Transactions on Information and Systems, vol. 101, no. 2, pp. 552-555.
23. Naka. F. R„ & Ward. W. W. (2000). Distant early warning line radars: The quest for automatic signal detection. Lincoln Laboratory Journal, vol. 12, no. 2, pp. 181-204.
24. Perov A.L (2003), Statisticheskaja teorija radiotehnicheskih system [Statistical theory of radio technical systems]. Moscow: Radiotchnika, 400p, (in Russian)
25. Grishin, U.P., Ipatov, V.P, and Kazarynov, U.M. (1990), Radiotehnicheskie sistemy [Radar systems], in Kazarynov, U.M. (ed.). Moscow: Vysshaja shkola, 496p, (in Russian)
26. Punt, M.. Bruijn, S. M„ Wittink, H„ & van Dieen, J. H. (2015). Effect of arm swing strategy on local dynamic stability of human gait. Gait posture, vol. 41, no. 2, pp. 504-509.
27. Ghaleb, A., Vignaud, L„ & Nicolas, J. M. (2008). Micro-Doppler analysis of pedestrians in 1SAR imaging. The IET Signal Processing, vol. 2, no. 3, pp. 301-311.
28. Garcia-Rubia, J. M., Kilic, O., Dang, V„ Nguyen, Q. M„ & Tran, N. (2014). Analysis of moving human micro-Doppler signature in forest environments. Progress In Electromagnetics Research, vol. 148, pp. 1-14.
29. AWRI243 76-Gllz to SI-GHz high-performance automotive MMIC evaluation module (2019). [online] Available at: http://www.ti.coni/tooI/AWR1243BOOST (Accessed 7 Jun.2019).
7TT
МЕТОД АВТОМАТИЧЕСКОГО РАСПОЗНАВАНИЯ ПЕШЕХОДОВ В ДОРОЖНОЙ СЦЕНЕ ПО СИГНАЛУ МИКРО-ДОПЛЕРА ДЛЯ РАДИОЛОКАЦИОННЫХ СИСТЕМ БЕСПИЛОТНОГО
АВТОТРАНСПОРТА
Плучевский Андрей Владимирович, АО "КОГНИТИВ", г. Москва, Россия; Томский государственный университет систем управления и радиоэлектроники (ТУСУР), г. Томск, Россия,
pluch.andry@gmail.com
Аннотация
Предложен метод распознавания пешехода по уникальной характеристике спектрограммы Доплера, измеренной с высоким разрешением. Задача распознавания решается в дорожной сцене на фоне движущихся автомобилей. Доплеровская спектрограмма рассматривается как двухмерное радиолокационное изображение. Учитывая особенности сигнала микро-Доплера пешехода, обработка двухмерного изображения сведена к одномерному пороговому. Алгоритм разработан для использования в режиме скользящего окна, при непрерывном поступлении информации о доплеровском спектре цели. Метод получен при аналитическом решении задачи и не требует применения подходов машинного обучения и нейронных сетей. Экспериментальное исследование проводились на двух типах объектов дорожной сцены: пешеходах и автомобилях. Показано, что метод способен отличить пешехода от автомобиля даже при сопоставимой мгновенной полосе Доплеровского спектра цели. Исследована вероятность правильного обнаружения цели разработанного метода в зависимости от отношения сигнал/шум и вероятности ложной тревоги. Это позволяет задать требование к радиолокационной система на этапе проектирования и оценить возможность использования метода в уже существующих радиолокационных комплексах. Метод подходит для применения как в области радиолокации, так и в области компьютерного зрения. Метод разрабатывался для применения в индустрии беспилотного автотранспорта и систем помощи водителю для распознавания пешеходов и принятия необходимых мер для предотвращения столкновений в непредвиденных ситуация.
Ключевые слова: микро-Доплер, преобразование Фурье, автоматическое распознавание целей, распознавание пешеходов, каденсная диаграмма, автомобильный радиолокатор.
Литература
1. Patole, S. M., Torlak, M., Wang, D., & Ali, M. (2017). Automotive radars: A review of signal processing techniques. IEEE Signal Processing Magazine, vol. 34, no. 2, pp. 22-35.
2. Hasch, J. (2015). Driving towards 2020: Automotive radar technology trends. The 2015 IEEE MTT-S International Conference on Microwaves for Intelligent Mobility (ICMIM), pp. 1-4.
3. Kim, J., Han, D. S., & Senouci, B. (2018). Radar and vision sensor fusion for object detection in autonomous vehicle surroundings. The 2018 Tenth International Conference on Ubiquitous and Future Networks (ICUFN), pp. 76-78.
4. Steinbaeck, J., Steger, C., Holweg, G., & Druml, N. (2017). Next generation radar sensors in automotive sensor fusion systems. The 2017 Sensor Data Fusion: Trends, Solutions, Applications (SDF), pp. 1-6.
5. Горелик АЛ., Барабашh Ю.Д.., Кривошеев О.В. (1990). Селекция и распознавание на основе локационной информации. M.: Радио и связь, 240 с.
6. Pluchevskiy A.V. (2019). Micro-Doppler feature extraction of pedestrian and automobile using two-dimensional fast Fourier transform. T-Comm, vol. 13, no.5, pp. 61-68.
7. Bartsch, A., Fitzek, F., & Rasshofer, R. H. (2012). Pedestrian recognition using automotive radar sensors. Advances in Radio Science, vol. 10, no. 2, pp. 45-55.
8. Nanzer, J. A., & Rogers, R. L. (2009). Bayesian classification of humans and vehicles using micro-Doppler signals from a scanning-beam radar. IEEE Microwave and Wireless Components Letters, vol. 19, no. 5, pp. 338-340.
9. Lee, S., Yoon, Y. J., Lee, J. E., & Kim, S. C. (2017). Human-vehicle classification using feature-based SVM in 77-GHz automotive FMCW radar. IET Radar, Sonar & Navigation, vol. 11, no. 10, pp. 1589-1596.
10. Mesloub, A., Abed-Meraim, K., & Belouchrani, A. (2017). Ground moving target classification based on micro-Doppler signature using novel spectral information features. The 2017 European Radar Conference (EURAD), pp. 255-258.
T-Comm ^м 13. #8-2019
11. Zenaldin, M., & Narayanan, R. M. (2016). Radar micro-Doppler based human activity classification for indoor and outdoor environments, Proc. SPIE, [online] vol. 9829, no. 98291. Available at: http://dx.doi.org/I0.lll7/l2.2228397 (Accessed 18 Feb.2019)
12. Cao, P., Xia, W., & Li, Y. (2018). Classification of Ground Targets Based on Radar Micro-Doppler Signatures Using Deep Learning and Conventional Supervised Learning Methods. Radioengineering, vol. 27, no. 3, pp. 835-845.
13. Erol, B., Amin, M. G., & Gurbuz, S. Z. (2018). Automatic data-driven frequency-warped cepstral feature design for micro-Doppler classification. IEEE Transactions on Aerospace and Electronic Systems, vol. 54, no. 4, pp. 1724-1738.
14. Tivive, F. H. C., Phung, S. L., & Bouzerdoum, A. (2015). Classification of micro-Doppler signatures of human motions using log-Gabor filters. IET Radar, Sonar & Navigation, vol. 9, no. 9, pp. 1188-1195.
15. Bjrklund, S., Petersson, H., & Hendeby, G. (2015). Features for micro-Doppler based activity classification. IET radar, sonar & navigation, vol. 9, no. 9, pp. II8I-II87.
16. Kim, Y., & Moon, T. (20I5). Human detection and activity classification based on micro-Doppler signatures using deep convolutional neural networks. IEEE geoscience and remote sensing letters, vol. I3, no. I, pp. 8-I2.
17. Erol, B., & G?rb?z, S. Z. (20I6). Hyperbolically-warped cepstral coefficients for improved micro-Doppler classification. The 20I6 IEEE Radar Conference (RadarConf), pp. I-6.
18. Bjrklund, S., & Rydell, J. (20I7). Micro-Doppler classification with boosting in perimeter protection. Belfast, UK. The International Conference on Radar Systems, pp. 23-26.
19. Gurbuz, S. Z., Erol, B., Cagliyan, B., & Tekeli, B. (20I5). Operational assessment and adaptive selection of micro-Doppler features. IET Radar, Sonar & Navigation, vol. 9, no. 9, pp. II96-I204.
20. Du, R., Fan, Y., & Wang, J. (2018). Pedestrian and Bicyclist Identification Through Micro Doppler Signature With Different Approaching Aspect Angles. IEEE Sensors Journal, vol. I8, no. 9, pp. 3827-3835.
21. Li, W., Xiong, B., & Kuang, G. (2017). Target classification and recognition based on micro-Doppler radar signatures. The 2017 Progress in Electromagnetics Research Symposium-Fall (PIERS-FALL), pp. I679-I684.
22. Fan, Y., Du, R., & Wang, J. (2018). Identification of Pedestrian and Bicyclist through Range Micro Doppler Signatures. IEICE Transactions on Information and Systems, vol. I0I, no. 2, pp. 552-555.
23. Naka, F. R., & Ward, W. W. (2000). Distant early warning line radars: The quest for automatic signal detection. Lincoln Laboratory Journal, vol. I2, no. 2, pp. I8I-204.
24. Перов А И. (2003). Статистическая теория радиотехнических систем. M.: Радиотехника, 400 с.
25. Гришин Ю. П., Ипатов В. П., Казаринов Ю. М. (I990). Радиотехнические системы. Под ред. Ю.М. Казаринова. М.: Высш. шк., 496 с.
26. Punt, M., Bruijn, S. M., Wittink, H., & van Dieеn, J. H. (20I5). Effect of arm swing strategy on local dynamic stability of human gait. Gait & posture, vol. 4I, no. 2, pp. 504-509.
27. Ghaleb, A., Vignaud, L., & Nicolas, J. M. (2008). Micro-Doppler analysis of pedestrians in ISAR imaging. The IET Signal Processing, vol. 2, no. 3, pp. 30I-3II.
28. Garcia-Rubia, J. M., Kilic, O., Dang, V., Nguyen, Q. M., & Tran, N. (20I4). Analysis of moving human micro-Doppler signature in forest environments. Progress In Electromagnetics Research, vol. I48, pp. I-I4.
29. AWRI243 76-GHz to 8I-GHz high-performance automotive MMIC evaluation module (20I9). [online] Availible at: http://www.ti.com/tool/AWRI243BOOST (Accessed 7 Jun.20I9)
Информация об авторе:
Плучевский Андрей Владимирович, Младший инженер разработчик, департамент радиолокации, AO "КОГНИТИВ", Москва, Россия;
аспирант, ассистент кафедры телекоммуникаций и основ радиотехники (ТОР), Томский государственный университет систем управления и радиоэлектроники (ТУСУР), г. Томск, Россия
( I л
