1. Introduction
Using millimeter-wave radar as a monitoring system has recently raised increased in various fields attributed to its high detection accuracy and weather resistance. In road traffic applications, accurate acquisition of moving target information such as pedestrians and vehicles on the road is one of traffic radar’s most basic and essential tasks. However, with the development of intelligent transportation systems, the timely reflection of road conditions and realization of traffic accident warnings become necessary, which puts higher requirements on the low processing delay of radar systems. Therefore, we start from the radar target detection method to reduce the system processing delay and improve the system’s real-time while ensuring high monitoring performance.
Generally, a number of approaches have been proposed to improve radar performance in traffic scenes, which can be mainly classified into two categories: (1) Around the signal processing direction, reduce or eliminate the interference signal as much as possible to ensure the purity of the target signal. (2) Design better detectors.
In the field of suppressing interference signals, cutting the interference-contaminated samples out of the signal is the most direct interference suppression method. Still, it also suppresses part of the valid signal of targets, resulting in less accurate reconstruction of the cutout samples of useful signals. To alleviate this effect, an iterative matrix-pencil (MP) method-based extrapolation for interference mitigation is proposed [
1], but the accuracy of the reconstructed signal will decrease with the increase of the proportion of the contaminated samples. To improve signal reconstruction accuracy, an approach based on the sparse and low-rank decomposition of the Hankel matrix is proposed [
2], however, iteration and best quality selection increase the complexity of the algorithm. For dealing with inter-radar interference, some new methods are proposed, such as designing a new orthogonal noise waveform [
3] or presenting an interference mitigation technique in the tunable Q-factor wavelet transform (TQWT) domain [
4]. In addition, some researchers have started from range and Doppler cell migration calibration to reduce signal distortion [
5].
Designing better detectors is another direction to improve the application ability of radar. Compared with signal interference suppression, although the improvement of radar detection performance is weak, the design of the detector is more straightforward. In particular, in most cases, it will not add additional links and computational complexity, which is more suitable for low-cost radar systems with low-performance processors. So this paper studies from the perspective of detector design. In the actual application of radar, the target always appears before the complex and changeable noise background, which is subject to changes in time and location, so the fixed threshold detection method cannot effectively distinguish the target from the background noise [
6]. Radar Constant False Alarm Rate(CFAR) detection, which adaptively sets the detection threshold by evaluating current clutter environments [
7,
8], is a crucial technology for separating targets and background. Thus, an appropriate CFAR detection algorithm can improve radar detection performance and provide accurate target information for radar data processing. It also emerged as a key research aspect in radar detection.
The earliest proposed CFAR detection algorithm is the Cell Average CFAR (CA-CFAR) [
9], which is based on the average local noise power level to detect targets. Design a reference window, take the window’s center unit as the candidate target, and use the amplitude accumulation value of other units in the window as the estimated value of background noise power. Whether the target exists is determined by judging the power value of the candidate target and the estimated noise power value. The CA-CFAR has poor detection performance under non-uniform noise and multi-target environments since interference signals or other targets within the reference window lead to background noise estimation error. Aiming at the problem of CA-CFAR detection performance degradation under multi-target or non-uniform noise, the maximum selection detection (GO-CFAR) [
10] and the minimum selection detection (SO-CFAR) [
11] are proposed. The GO-CFAR can maintain good false alarm control performance in a clutter edge environment, but ”target masking” in a multi-target environment will appear. The SO-CFAR has a good multi-target resolution ability, but its false alarm control ability is weak. Ordered Statistical CFAR (OS-CFAR) [
12] is another typical CFAR detection method derived from the median filter concept in digital image processing. It arranges the sample units in the reference window according to the power value and selects one of the sample power values as the decision threshold. Compared with the detection performance of CA-CFAR, OS-CFAR has strong robustness to multi-target detection, but it has a high resulting CFAR loss. Some new methods combining OS and CA are proposed. Take the trimmed-mean detector (TM-CFAR) [
13] algorithm as an example. The sample points in the reference window are sorted by amplitude value like OS-CFAR. Then remove part of the sample points with the largest and smallest amplitudes. Finally, the remaining sample points are averaged as the estimated value of background noise.
Now, new CFAR detection algorithms have been proposed based on traditional methods for different application requirements. Some scholars are dedicated to studying the multi-target detection capabilities of the detector. For example, the OSCA-CFAR algorithm is proposed, which combines the CA-CFAR algorithm and the OS-CFAR algorithm, to improve the multi-target detection capability of millimeter-wave radars, and gives algorithm performance simulation under an ideal noise environment [
14]. Some researchers focus on improving the detection performance of the detector in a complex cluster environment. A Comp-CFAR method is proposed according to the central limit theorem and the logarithmic compression principle of the signal for targets detection in the clutter with long smearing effect characteristics [
15]. A CFAR detector based on zlog(z) is used to reduce the false alarm rate in Weibull clutter [
16]. Some other scholars have proposed a new method combined with machine learning [
17]. Train a feedforward artificial neural network (ANN) on the data set containing the input and output of the CFAR sliding window to improve the efficiency of CFAR detection. However, this way increases the complexity of the algorithm.
Table 1 briefly summarizes the CFAR detection algorithms mentioned above. All current CFAR algorithms achieve target detection by designing a reference window and processing the data in the window. The design of the reference window can effectively reduce the level of clutter interference, but there are two drawbacks. One is that sliding the window reduces the efficiency of detecting objects. Especially in Linear Frequency Modulated Continuous Wave (LFMCW) radar systems, the input to the CFAR detector is a two-dimensional Range–Doppler Matrix (RDM). Although the principle of the CFAR algorithm has not changed, the sliding of the reference window has changed from a one-dimensional sliding search to a two-dimensional sliding search, which dramatically reduces the real-time performance of radar target detection. The other drawback is that the window limits the sample points for background noise estimation. Theoretically, the more sample points are used, the more accurate the estimated noise power value and the higher the CFAR detection accuracy. The design and sliding of the reference window limit the efficiency of radar target detection.
Fortunately, compared with the complex background environment such as the sea surface, the background noise in the traffic road is relatively simple. Especially when radar monitors moving targets, the effect of background noise data on detection can be eliminated using the de-zeroing Doppler method. Combining the non-ideal target motion (vehicle turning, braking and lane changing), the irregularity of the target reflection area, and the independence of the frame data, we consider that the moving target echoes in the traffic road obey the ideal Swerling II model. Based on the above analysis, a reference window designed for complex noise and interference reduction may not be necessary for road traffic monitoring applications where the background noise is relatively clean. Instead, the sliding window increases the algorithm’s time complexity and reduces the efficiency of radar monitoring. To meet the requirement of low processing latency for traffic monitoring radar systems, we propose a CFAR algorithm based on the Monte Carlo to improve radar detection efficiency and sensitivity to moving targets in the traffic environment. Compared with the traditional algorithm, the algorithm has higher detection sensitivity and, more importantly, does not require the design and sliding of the reference window, which dramatically reduces the algorithm time complexity and improves the detection speed and efficiency. The contributions of this paper are summarized as follows: First, in the proposed CFAR detection algorithm, the entire RDM matrix area is randomly sampled to obtain the sample points of the background noise power estimation at the current moment. The principle is converting the Monte Carlo simulation principle, which gets properties of unknown quantities by independently repeating experiments, into a random sampling of the RDM matrix domain. This way breaks through the limitation of the reference window and can increase a large number of sample points for background noise estimation. Second, an interference point filtering method improves the accuracy of background noise estimation. Sort and filter the sample points according to the amplitude value of the sample points to remove possible target points or interference points. Then, the background noise estimated power value is obtained by the mean method. Finally, the target points in the RDM matrix are extracted by the background noise estimate. In addition, the parameter setting method of the algorithm is given, i.e., according to the current physical platform and application environment, algorithm parameter values are obtained through repeated statistics. The configuration parameter process only needs to be executed once when the radar system environment is unchanged. Compared with the conventional method, simulations and practical experiments show that the method has higher detection sensitivity, higher detection accuracy, and lower detection latency, which improve radar detection efficiency in traffic surveillance.
The structure of this paper is organized as follows: The
Section 2 introduces the characteristics of traffic scenes, the principle of radar detection targets, and the characteristics of radar background noise in traffic scenes. Then, in
Section 3, we give the principle and model of the new CFAR detection algorithm and give the process of obtaining relevant parameters. In
Section 4, we analyze the performance of the new CFAR detection algorithm through simulation and real experiments. Finally,
Section 5 summarizes the conclusions of this paper.