A track initiation method based on rule and FasterRCNN model

Abstract

The present invention discloses a track initiation method based on a combination of rules and FasterRCNN models, which includes the following steps: obtaining radar echo information at N consecutive moments; generating a radar echo image according to the radar echo information; detecting the radar echo image using FasterRCNN to obtain a number of candidate areas; performing angle constraint judgment and filter prediction judgment on the echo point information in each candidate area in turn to obtain track initiation information; generating a radar echo image according to the radar echo information, and then detecting the radar echo image in combination with FasterRCNN, and utilizing the image processing capability of FasterRCNN to effectively extract candidate frames where tracks may exist; finally screening the candidate frames by combining angle constraint and filter prediction judgment, taking into account the advantages of sequential processing and batch processing methods, thereby reducing the false start rate and improving the performance of target track initiation.

Description

Technical Field

[0001] The present invention belongs to the technical field of target tracking, and in particular relates to a track initiation method based on a combination of rules and a FasterRCNN model. Background Art

[0002] Target tracking is the process of modeling, estimating, and tracking the motion of an object using various observational and computational methods. Since its conception, its research findings have been widely applied in both military and civilian fields. Military applications include missile interception, aircraft detection and tracking, early warning and penetration, and real-time battlefield monitoring. Civilian applications include ground-based human tracking, traffic regulation, air traffic control, maritime vessel monitoring, the currently popular unmanned vehicle (AV) system, and computer vision applications such as facial recognition and gesture tracking. Radar, with its all-weather operation, has become a key sensor in strategic defense systems, and target tracking based on radar echo points has also become a highly sought-after technical tool.

[0003] Track initiation is the first step in target tracking. It refers to the process of establishing the existence of a track from the radar's detected echo signal. It is the primary prerequisite and key task for target tracking. Failure to initiate a track correctly not only affects the determination of the target number but may also result in target loss, significantly impacting track maintenance and subsequent situation assessment tasks. Classic track initiation methods can be roughly divided into two categories: sequential processing methods, represented by intuitive and logical methods, and batch processing methods, represented by Hough transforms and their improved methods.

[0004] The main idea of ​​the sequential processing method is to process echoes according to the time series, taking all the echoes of the first cycle as the track head, combining them with the following echoes, setting relevant speed, acceleration, angle, etc. as constraints and establishing gates to judge and limit the generation of tracks, and integrating Kalman filtering and the logic of determining the start, which can be used for the entire tracking process and is relatively simple and intuitive.

[0005] The batch processing method superimposes echo data from multiple cycles within the radar field of view and uses Hough line detection from image processing to identify possible tracks in the spatial dimension. This avoids the combinatorial explosion problem encountered with sequential processing. Both methods have their advantages and disadvantages, but both strive to strike a balance between correctly initiated tracks and falsely initiated tracks, achieving faster and more accurate track initiation.

[0006] As radar detection environments become increasingly complex, current track initiation methods rely heavily on empirical knowledge in practical applications. This leads to weak generalization and poor adaptability to environmental changes, making them prone to problems such as track fragmentation and false tracks. For example, in sequential processing methods, because each echo point at the current moment must determine whether it satisfies the set constraints with the echo points at the previous and subsequent moments, dense clutter environments often lead to combinatorial explosion. Furthermore, since the hyperparameter settings in the prior constraints may not be consistent with real-world scenarios, a large number of false tracks can be initiated. The same is true for batch processing methods. Although superimposing echoes at multiple moments can avoid combinatorial explosion in dense clutter environments, the algorithm's principle is to use the linear relationship between points in the graph to initiate tracks. This is also difficult to implement in the presence of measurement noise due to the difficulty in selecting parameters for partitioning the space, often leading to the appearance of many false tracks. Furthermore, methods based on Hough transforms often suffer from track clustering. Summary of the Invention

[0007] The purpose of the present invention is to provide a track initiation method based on the combination of rules and FasterRCNN model, which takes into account the advantages of sequential processing and batch processing methods, reduces the false start rate, and improves the performance of target track initiation.

[0008] The present invention adopts the following technical solution: a track initiation method based on a combination of rules and FasterRCNN model, comprising the following steps:

[0009] Obtain radar echo information at N consecutive moments;

[0010] generating a radar echo image according to radar echo information;

[0011] Use FasterRCNN to detect the radar echo image and obtain several candidate regions;

[0012] The echo point information in each candidate area is subjected to angle constraint judgment and filtering prediction judgment in turn to obtain the track starting information.

[0013] Furthermore, generating a radar echo image according to the radar echo information includes:

[0014] Determine the side length of the radar echo image according to the radar field of view detection range;

[0015] The radar echo information is compressed based on the side length and drawn into the radar echo image; wherein the radar echo information at different times has different colors.

[0016] Furthermore, performing angle constraint judgment on the echo point information in each candidate area includes:

[0017] Generate a first vector based on the pixel corresponding to the echo point at time n and the pixel corresponding to the echo point at time n+1, where n∈N, n≠1 and n≠N;

[0018] Generate a second vector based on the pixel corresponding to the echo point at time n and the pixel corresponding to the echo point at time n+2;

[0019] Calculate the angle between the first vector and the second vector;

[0020] When the included angle is less than or equal to the angle constraint threshold, continue to perform filtering prediction judgment.

[0021] Furthermore, when the included angle is greater than the angle constraint threshold, the echo point in the candidate area is deleted.

[0022] Furthermore, filtering and predicting the echo point information in each candidate area includes:

[0023] Taking the set of pixels corresponding to the non-first and last echo points in N consecutive moments as the known quantity, forward or backward Kalman filtering is performed to obtain the area where the pixels corresponding to the predicted first and last echo points are located;

[0024] When the area where the pixels corresponding to the first and last echo points are predicted to be located contains the pixels corresponding to the first and last echo points, the set of pixels corresponding to the echo points at N consecutive moments is taken as the starting point of the track.

[0025] Furthermore, when the region where the pixels corresponding to the first and last echo points are predicted to be located does not include the pixels corresponding to the first and last echo points, the echo points in the candidate region are deleted.

[0026] Furthermore, performing forward or backward Kalman filtering includes:

[0027] The average value of the pixel points corresponding to the non-first and last echo points in N consecutive moments is used as the initial intermediate moment state, and the echo at the previous moment and the echo at the next moment are selected from the pixel points corresponding to the non-first and last echo points in N consecutive moments as the correlation measurement;

[0028] Perform forward filtering based on the initial intermediate state and the previous moment echo to obtain the pixel point corresponding to the initial moment state prediction point, and determine the initial moment pixel point area based on the pixel point corresponding to the initial moment state prediction point;

[0029] Backward filtering is performed based on the initial intermediate state and the echo at the next moment to obtain the pixel points corresponding to the final moment state prediction point, and the final moment pixel point area is determined based on the pixel points corresponding to the final moment state prediction point.

[0030] Furthermore, determining the pixel point area at the initial moment based on the pixel point corresponding to the state prediction point at the initial moment includes:

[0031] The pixel point area at the initial moment is generated with the pixel point corresponding to the state prediction point at the initial moment as the center.

[0032] Furthermore, when training FasterRCNN:

[0033] In the radar echo image used as a training sample, the track start labeling box is larger than the track trajectory.

[0034] Another technical solution of the present invention: a track initiation device based on the combination of rules and FasterRCNN model, including a memory, a processor and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the above-mentioned track initiation method based on the combination of rules and FasterRCNN model is implemented.

[0035] The beneficial effects of the present invention are as follows: the present invention generates a radar echo image based on radar echo information, and then detects the radar echo image in combination with FasterRCNN. The image processing capability of FasterRCNN can be used to effectively extract candidate frames where tracks may exist. Finally, the candidate frames are screened by combining angle constraints and filtering prediction judgment, taking into account the advantages of sequential processing and batch processing methods, thereby reducing the false start rate and improving the performance of target track initiation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of a flow chart of a track initiation method based on a combination of rules and FasterRCNN model according to an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of Hough transform parameter selection in an embodiment of the present invention;

[0038] Figure 3 This is a comparison diagram of the track start in simulation scenario 1 in the embodiment for verifying the present invention;

[0039] Figure 4 This is a comparison chart of the starting rate curve of simulation scenario 1 in the embodiment for verifying the present invention;

[0040] Figure 5 This is a comparison chart of the initiation rate curves of simulation scenario 2 in the embodiment for verifying the present invention. DETAILED DESCRIPTION

[0041] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] With the continuous development of machine learning and deep learning, data-driven methods for track initiation have emerged in recent years. Due to the limited information content and temporal nature of radar tracks, these methods often cannot exist independently. The general approach is to first use intuitive methods or logical initiation to achieve a rough track initiation. Then, corresponding feature data is constructed from these tracks and input into traditional machine learning classifiers or deep learning networks to classify and distinguish true and false tracks.

[0043] Therefore, in order to address some of the shortcomings of the existing methods, the present invention designs a track initiation method based on the combination of rules and FasterRCNN model, which takes into account the advantages of sequential processing and batch processing methods, reduces the false start rate, and improves the performance of target track initiation.

[0044] Track initiation is the primary task in radar point target tracking. Traditional methods often require sufficient prior information and accurate prior parameters. Batch processing methods such as intuitive and logical methods, as well as those based on Hough transforms, often lead to problems such as track fragmentation and multiple false tracks. Therefore, a data-driven approach is considered, combining the FasterRCNN deep detection network and rule constraints to complete track initiation. Radar echoes at multiple times are converted into images of different colors and input into the deep detection network to complete the learning and detection task. Candidate boxes of possible targets are drawn in the image. Linear angle constraints are set around this candidate area and extrapolated to the corresponding echo points to complete track initiation.

[0045] Specifically, the present invention discloses a track initiation method based on a combination of rules and FasterRCNN model, such as Figure 1 As shown in FIG, the method includes the following steps: obtaining radar echo information at N consecutive moments; generating a radar echo image according to the radar echo information; detecting the radar echo image using FasterRCNN to obtain several candidate areas; and performing angle constraint judgment and filtering prediction judgment on the echo point information in each candidate area in turn to obtain the track starting information.

[0046] The present invention generates a radar echo image based on radar echo information, and then detects the radar echo image in combination with FasterRCNN. The image processing capability of FasterRCNN can be used to effectively extract candidate frames where tracks may exist. Finally, the candidate frames are screened by combining angle constraints and filtering prediction judgment, taking into account the advantages of sequential processing and batch processing methods, thereby reducing the false start rate and improving the performance of target track initiation.

[0047] More specifically, the radar echo is first converted into image information. During this process, the side length of the radar echo image is determined based on the radar's field of view. Based on the side length, the radar echo information is compressed and mapped onto the radar echo image. Radar echo information at different times has different colors.

[0048] As a specific implementation, the maximum value of the radar's field of view is used as the image boundary. Limited by the data input format of the deep neural network and the limitations of computer hardware, the image is proportionally reduced to an image with a side length ranging from 10 to 1000 pixels. Within this range, the target's speed and distance-related parameters are compressed accordingly. The echo data at each moment is accumulated and plotted on the graph and assigned different pixel values. In this way, the echo points at different times are reflected in different colors in the graph to represent information in the time dimension. The shape and distribution of the overall echo accumulation in the graph represent the spatial information of the motion. This constructed image enables the deep detection network to fully utilize the echo data and discover potential candidate areas where tracks may exist.

[0049] Furthermore, the image is input into the deep detection network FasterRCNN. In the field of deep learning and image detection, FasterRCNN is already a classic and mature two-stage detection network, which has been widely recognized and applied in the industry. The present invention uses it as a detector, and its main purpose is to extract candidate areas where tracks may exist. After the image is input into the network, it first passes through the residual network (ResNet) and the feature pyramid network (FPN, Feature Pyramid Networks) to extract image features and obtain the feature map of the image. ResNet is a deep network composed of multiple residual blocks. ResNet50 is used here, that is, the network has 50 layers, all of which are composed of convolution, pooling and relu activation functions. FPN is a network structure strategy that fuses multi-scale features. It fuses feature maps at different scales to enhance the extraction of image features. The extracted feature map is then fed to the Region Proposal Network (RPN), which generates multiple anchor boxes with fixed aspect ratios based on the receptive field area of ​​the original image corresponding to each point in the feature map. Multiple convolutional layers are then used to determine whether the box contains the target. During the training phase, the candidate box is compared with the true box to complete the regression task, resulting in many candidate proposal regions that contain the target and are close to the true label box. Finally, the Region of Interest (RoI pooling) layer is used to extract features from the candidate proposal regions of different sizes and the feature maps extracted by convolution. After outputting fixed-size features, the fully connected network is connected to complete the classification of the target and correct the outer box of the candidate proposal region. At this point, the entire FasterRCNN network completes the target detection task.

[0050] To train the aforementioned deep neural network, a large number of training samples are required. When constructing these training samples, the target begins moving from any position and direction within the field of view. The simulated radar echo is used as input, converted into an image, and the target's initial and final moments are used as coordinate references for the label box. To fully encompass the target within the label box, the label box's boundaries are expanded by two pixels around the target. This means that in the radar echo image used as the training sample, the track start label box is larger than the track trajectory.

[0051] That is, assuming that the initial coordinates of the target are (x start ,y start ), the target coordinates detected in the last frame are (x end ,y end ), then the coordinates of the upper left point of the label box in the figure (x label1 ,y label1 ) and the coordinates of the lower right point (x label2 ,ylabel2 ) is represented as follows:

[0052] (x label1 ,y label1 )=(min(x start ,x end )-2,min(y start ,y end )-2),

[0053] (x label2 ,y label2 )=(max(x start ,x end )+2,max(y start ,y end )+2),

[0054] In this embodiment, there is only one target category, that is, the network only needs to complete the binary classification task of background and target, and all target category labels are the same. The loss function of the entire network is a weighted combination of the cross entropy classification loss function and the smoothed L1 regression loss.

[0055] Next, we use the trained deep learning network to extract candidate regions from the echo image. The trained FasterRCNN proposes candidate regions in the echo image that may contain flight paths, distinguishing them from the background. Given the inherent errors introduced by the network, which may not fully encompass the target area, we set a hyperparameter σ to expand the generated candidate region boundaries.

[0056] In an embodiment of the present invention, performing angle constraint judgment on the echo point information in each candidate area includes: generating a first vector based on the pixel point corresponding to the echo point at time n and the pixel point corresponding to the echo point at time n+1, where n∈N, n≠1 and n≠N; generating a second vector based on the pixel point corresponding to the echo point at time n and the pixel point corresponding to the echo point at time n+2; calculating the angle between the first vector and the second vector; and continuing to perform filtering prediction judgment when the angle is less than or equal to the angle constraint threshold.

[0057] Specifically, in the expanded candidate frame area, find the pixels at the second, third, and fourth moments (cyan, yellow, and green) by color, establish a vector relationship between each pixel at the second moment and the pixel at the third and fourth moments, and calculate the angle θ between the position vectors at the second, third, and fourth moments with the pixel at the second moment as the origin, and set the angle constraint threshold θ gate , if θ≤θ gate , that is, it is considered that there are three echo points at different times that satisfy the straight line segment relationship under the measurement noise of the current environment, and there is a possibility of starting the track. Otherwise, the echo points in the area are judged as clutter, and the echo points in the candidate area are deleted.

[0058] In another verification method, filtering and predicting the echo point information in each candidate area includes: using the set of pixels corresponding to non-first and last echo points in N consecutive moments as a known quantity, performing forward or backward Kalman filtering to obtain the area where the pixels corresponding to the predicted first and last echo points are located; when the area where the pixels corresponding to the predicted first and last echo points are located contains the pixels corresponding to the first and last echo points, the set of pixels corresponding to the echo points in N consecutive moments is used as the starting point of the track. Similarly, when the area where the pixels corresponding to the predicted first and last echo points are located does not contain the pixels corresponding to the first and last echo points, the echo point in the candidate area is deleted.

[0059] In other words, a linear extrapolation is performed on the echo points that meet the above requirements within the candidate area to determine whether the echo at the starting and last moments falls within the gate to complete the track initiation. There may be more than one echo point within the candidate area that meets the angle constraint, and the track initiation requires echoes at at least four of the five moments to meet the starting conditions. Therefore, the initial or last moment is required to complete the judgment.

[0060] In one embodiment, the average value of the pixel points corresponding to the non-first and last echo points in N consecutive moments is used as the initial intermediate moment state, and the echo of the previous moment and the echo of the next moment are selected as the associated measurement from the pixel points corresponding to the non-first and last echo points in N consecutive moments; forward filtering is performed based on the initial intermediate moment state and the echo of the previous moment to obtain the pixel points corresponding to the initial moment state prediction point, and the initial moment pixel point area is determined based on the pixel points corresponding to the initial moment state prediction point, that is, the initial moment pixel point area is generated with the pixel point corresponding to the initial moment state prediction point as the center. Backward filtering is performed based on the initial intermediate moment state and the echo of the next moment to obtain the pixel points corresponding to the final moment state prediction point, and the final moment pixel point area is determined based on the pixel points corresponding to the final moment state prediction point, that is, the final moment pixel point area is generated with the pixel point corresponding to the final moment state prediction point as the center.

[0061] More specifically, the coordinate average of the three echo points that are associated with each other in the candidate area and meet the angle constraint is used as the initial state position, and the average speed between these three moments is used as the initial state speed. That is, for a three-moment echo (x2, y2)(x3, y3)(x4, y4) associated with each other in the candidate box, the initial state x is constructed. mid =[x mid ,vx mid ,y mid ,vy mid ], then it is expressed as follows, where T is the sampling period.

[0062]

[0063]

[0064] The uniform linear motion (CV) model is used as the dynamic evolution model, and the second and fourth moment echoes are used as the initial state x mid The next moment of the associated echo, forward and backward Kalman filtering, and extrapolation prediction to obtain the initial moment prediction state and the predicted state at the last moment and their information covariance matrix and

[0065] Taking the echo judgment at the last moment as an example, the position coordinates of the predicted state at the last moment are As the center, set the square selection area threshold β gate , in the square area of ​​the entire image First, find out whether there is a pixel that represents the color of the last moment. If it exists, use it as the echo of the last moment to determine whether it falls into the center. And the covariance is If all conditions are met, the point is considered to be the last associated echo point. The same method is used to confirm the associated echo at the initial moment. At this point, the track initiation method described in this embodiment is completed.

[0066] Verification example:

[0067] A 2D radar clutter simulation environment is set up with a detection range of 0 to 5000 m in plane coordinates and a scanning period of T = 1 second. During the track initiation phase, the target is assumed to be in uniform linear motion with a velocity range of 200 to 250 m / s. Five targets are set, and their initial position coordinates are randomly generated within the two-dimensional plane. The radar detects for a total of five cycles, and the number of clutter waves in each cycle follows a Poisson distribution. By setting different clutter densities, the Poisson distribution parameter λ corresponding to the number of clutter waves per cycle within the radar detection range is obtained. The number of clutter waves generated per cycle is determined from this Poisson distribution, and these clutter waves are randomly distributed uniformly across the radar field of view. Next, the proposed method is compared with classic track initiation methods such as the 3 / 4 logic method and the Hough transform under varying measurement noise and clutter densities to verify its effectiveness.

[0068] The correct starting rate P of 1000 Monte Carlo simulations d and the false start rate P f As an indicator, its formula is as follows.

[0069]

[0070]

[0071] Among them, I it The indicator function representing the start of the i-th target in the i-th Monte Carlo simulation, that is, its value is 1 if the start is successful, otherwise it is 0. MC is the number of Monte Carlo simulations, and N is the number of real targets. i is the number of tracks at the start of the i-th Monte Carlo simulation, and f i It represents the number of false start tracks in this simulation.

[0072] Simulation scenario 1 measures noise in a two-dimensional plane at 10m, and the clutter density changes from 10 -5 ~10 -4 pcs / m 2 Methods compared with the present invention include the M / N logic method, the Hough transform method, and the modified Hough transform method. Considering that traditional initiation methods have high requirements for prior parameters, the parameter values ​​of each method are described in detail below.

[0073] The M / N logic method takes M=3, N=4 for a better initial effect, so the logic method described in this embodiment is the 3 / 4 method. In this setting, the maximum speed in the 3 / 4 logic method is 500m / s, the minimum speed is -500m / s, the initial track head distance constraint is 500m, the angle constraint is set to 0.01rad, and the probability of falling into the associated elliptical wave gate is selected as P G =0.9997.

[0074] The effectiveness of the standard Hough transform method depends mainly on the angle division and distance division of the parameter space and the voting threshold within the parameter space grid, but there is no reasonable and appropriate selection standard for this. Therefore, for the Hough transform-based method, multiple parameters are selected and parameters with relatively high accuracy and low false start rate are selected, such as Figure 2 As shown in the figure, the standard Hough transform is fixed at a voting threshold of 4 when the clutter density is 1*10 -5 pcs / m 2 The parameter selection diagram below. In summary, considering the difficulty of parameter selection based on Hough transform method, in the standard Hough transform parameter selection, the fixed angle is divided into Δθ=π / 120 and the voting threshold in the parameter space grid is 4, and N is flexibly adjusted. ρ To obtain the optimal distance division Δ ρ=R max / N ρ .

[0075] The modified Hough transform has a fixed angle division of Δθ=π / 150. It does not have a grid voting threshold, but has an additional threshold θ0 for judging the angle of the zero intersection, which is set to θ0=Δθ, and N is flexibly adjusted. ρ To obtain the optimal distance partition Δρ=R max / N ρ .

[0076] In the method proposed by the present invention, a more stringent constraint threshold is selected considering that the measurement noise is small. The candidate frame expansion parameter σ is set to 5, and the angle constraint threshold within the frame is θ gate =0.35rad, the extrapolated pixel selection threshold is set to β gate =15. The clutter density is 6*10 -5 Taking the scenario of 1000 square meters per square meter as an example, the results of comparing several starting methods are as follows Figure 3 shown.

[0077] As can be seen from the comparison diagram, the method described in this article can completely initiate the tracks of the five targets without any extra false tracks. The Hough transform-based initiation method not only produces false tracks, but also has the problem of track clustering inherent to the Hough transform. Although the modified Hough transform is slightly better than the standard Hough transform, it still has the problems of this type of method due to the lack of time series information and speed and other rule constraints. The 3 / 4 logic method not only initiates the real track but also introduces some other false tracks. The result comparison diagram clearly reflects the limitations of the traditional method. Whether it is sequential processing or batch processing, it has a strong dependence on the prior parameters. Only when the selection is appropriate can satisfactory results be obtained. In addition, the method based on Hough transform, which converts to the parameter space to find the intersection point to detect the straight line segment, is accompanied by the disadvantage of track clustering, which is difficult to overcome. Because of the selection of the candidate box area and the coexistence of rule constraints, the method of the present invention suppresses false tracks and has a better starting effect than the traditional method.

[0078] The comparison curve of the correct start rate and false start rate under simulation scenario 1 is as follows Figure 4 As shown in the figure, the correct initiation rate of each initiation algorithm decreases as clutter density increases. The method of the present invention exhibits the highest correct initiation rate, showing a significant advantage over the 3 / 4 logic method, the Hough variation method, and the modified Hough variation method. In the false initiation rate curve, the false tracks of each traditional method increase significantly with increasing clutter density, while the method of the present invention performs more stably. Its advantages become more prominent as clutter density increases, demonstrating its superior ability to suppress false initiation rates.

[0079] Note that in low-density environments, the false start rate of the logical method is slightly lower than that of the method of the present invention. This is because, for the logical start of calculating all echo points, under this clutter density and given parameters, the sequential processing method can better meet the requirements for distinguishing true target points from false echo points. As the clutter density increases and the number of echoes increases, the requirements for prior parameters become increasingly stringent, and the advantages of the method of the present invention are revealed. However, accurate prior information is often difficult to obtain. In simulation scenarios that are closer to reality, compared to the 3 / 4 logical method and the Hough transform and modified Hough transform, which are also batch processing methods, the method of the present invention considers the information at different times in the echo graph to represent different colors and uses this to propose possible candidate boxes. This is more reliable than simply detecting straight line segments of indifferent echo points in the graph. Corresponding constraints are set in the candidate boxes, greatly suppressing false tracks formed by irrelevant echo points caused by the start.

[0080] Simulation Scenario 2 was conducted under a 2D measurement noise of 50m. Unlike Simulation Scenario 1, the increased measurement noise made it more difficult for the track to appear "straight." The Hough transform performed poorly under the low measurement noise of Simulation Scenario 1, so only the modified Hough transform was compared here. Hyperparameters for each method were also set to be relatively relaxed to prevent missed detections.

[0081] The 3 / 4 logic method sets the maximum speed to 500m / s, the minimum speed to -500m / s, the initial track head distance gate to 500m, and the angle gate to 0.01rad. The modified Hough transform fixed angle division is Δθ=π / 150, the threshold of the zero crossing angle is set to θ0=Δθ, and N is flexibly adjusted. ρ To obtain the optimal distance partition Δρ=R max / N ρ In the method proposed by the present invention, the angle constraint threshold within the candidate frame is set to θ gate =1rad, the extrapolated pixel selection threshold is set to β gate =20.

[0082] The track initiation rate curves of several methods are shown in the figure below: Figure 5 As shown. Figure 5In the comparison curve of the correct start rate and the false start rate, it can be seen that the method of the present invention still maintains the same advantage as the simulation scenario 1 when compared with the modified Hough transform and the 3 / 4 logic method. The correct start rate of the three methods has decreased compared with the simulation scenario 1, which is caused by the increase in measurement noise. The phenomenon that the Hough transform method exceeds the method of the present invention in the correct start rate can also explain the contradiction between the correct start rate and the false start rate. The Hough transform relies on the accumulation of echo points in the graph to find a straight line. When the measurement noise increases, the number of targets that meet the straight line requirements becomes smaller and smaller. At this time, more tracks can only be started by adjusting the prior hyperparameters. In this process, although there is a certain probability of starting a real target, it is more likely to bring about interference from false tracks. The method of the present invention is not easily affected by this when using a deep detection network to locate the area where the target may exist. It is only necessary to select a small number of hyperparameters when establishing rule constraints on the echo points in the candidate frame, that is, the demand for prior parameters will be smaller, which can give play to the advantage of using data-driven to replace part of the prior information.

[0083] In summary, current track initiation methods rely heavily on empirical knowledge in practical applications, have poor adaptability to environmental changes, and are prone to problems such as missed track detection and false tracks. The track initiation method proposed in this paper considers converting radar cumulative echoes into image information for processing. By combining data-driven and rule-based approaches, it comprehensively considers the target's spatial morphological and temporal information, leveraging the advantages of both sequential and batch processing methods to significantly reduce the rate of false track initiation. Compared to traditional track initiation methods, only rough hyperparameter adjustments based on clutter density and measurement noise are required, simplifying the hyperparameter selection and tuning process of traditional methods. Replacing prior information with large amounts of radar data overcomes the difficulty in obtaining precise prior parameters.

[0084] The radar echo is converted into image information that combines time and space for processing. With the help of the good detection, positioning and recognition capabilities of the deep detection network, candidate boxes that may contain tracks in the echo image are extracted, and corresponding starting rules are set. By combining radar data-driven and partial prior rules, the traditional method's demand for precise prior parameters is weakened, and the spatial and temporal information of the target is comprehensively considered, which greatly suppresses the false track initiation rate.

[0085] The present invention also discloses a track initiation device based on the combination of rules and FasterRCNN model, including a memory, a processor and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the above-mentioned track initiation method based on the combination of rules and FasterRCNN model.

[0086] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices are based on the same concept as the embodiment of the method of the present invention. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0087] The device can be a computing device such as a desktop computer, laptop, PDA, radar, or cloud server. The device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the device may include more or fewer components, or a combination of certain components, or different components, such as input / output devices, network access devices, etc.

[0088] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0089] In some embodiments, the memory may be an internal storage unit of the extraction device, such as a hard disk or memory of the extraction device. In other embodiments, the memory may also be an external storage device of the extraction device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card, etc. equipped on the extraction device. Furthermore, the memory may also include both an internal storage unit of the extraction device and an external storage device. The memory is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory may also be used to temporarily store data that has been output or is to be output.

Claims

1. A track initiation method based on rules combined with FasterRCNN model, characterized in that: The following steps are involved: Obtain radar echo information at N consecutive moments; generating a radar echo image according to the radar echo information; Using FasterRCNN to detect the radar echo image, several candidate regions are obtained; Angle constraint judgment and filtering prediction judgment are performed on the echo point information in each candidate area in turn to obtain track starting information.

2. A track initiation method based on rule combined with FasterRCNN model as claimed in claim 1, characterized in that: Generating a radar echo image according to the radar echo information includes: Determine the side length of the radar echo image according to the radar field of view detection range; The radar echo information is compressed based on the side length and drawn into the radar echo image; wherein the radar echo information at different times has different colors.

3. A track initiation method based on rule combined with FasterRCNN model as claimed in claim 2, characterized in that: Performing angle constraint judgment on the echo point information in each candidate area includes: Generate a first vector based on the pixel corresponding to the echo point at time n and the pixel corresponding to the echo point at time n+1, where n∈N, n≠1 and n≠N; Generate a second vector based on the pixel corresponding to the echo point at time n and the pixel corresponding to the echo point at time n+2; Calculating the angle between the first vector and the second vector; When the angle is less than or equal to the angle constraint threshold, the filtering prediction judgment is continued.

4. A track initiation method based on rule combined with FasterRCNN model as claimed in claim 3, characterized in that: When the included angle is greater than the angle constraint threshold, the echo points in the candidate area are deleted.

5. A track initiation method based on a combination of rules and FasterRCNN model according to any one of claims 2 to 4, characterized in that: Performing filtering prediction and judgment on the echo point information in each candidate area includes: Taking the set of pixels corresponding to the non-first and last echo points in N consecutive moments as the known quantity, forward or backward Kalman filtering is performed to obtain the area where the pixels corresponding to the predicted first and last echo points are located; When the area where the pixels corresponding to the first and last echo points are predicted to be located contains the pixels corresponding to the first and last echo points, the set of pixels corresponding to the echo points at N consecutive moments is taken as the starting point of the track.

6. A track initiation method based on rule combined with FasterRCNN model as claimed in claim 5, characterized in that: When the region where the pixel points corresponding to the first and last echo points are predicted to be located does not include the pixel points corresponding to the first and last echo points, the echo points in the candidate region are deleted.

7. A track initiation method based on rule combined with FasterRCNN model as claimed in claim 6, characterized in that: Performing forward or backward Kalman filtering involves: The average value of the pixel points corresponding to the non-first and last echo points in N consecutive moments is used as the initial intermediate moment state, and the echo at the previous moment and the echo at the next moment are selected from the pixel points corresponding to the non-first and last echo points in N consecutive moments as the correlation measurement; Perform forward filtering based on the initial intermediate state and the previous moment echo to obtain the pixel point corresponding to the initial moment state prediction point, and determine the initial moment pixel point area based on the pixel point corresponding to the initial moment state prediction point; Backward filtering is performed based on the initial intermediate state and the echo at the next moment to obtain the pixel points corresponding to the final moment state prediction point, and the final moment pixel point area is determined based on the pixel points corresponding to the final moment state prediction point.

8. A track initiation method based on rule combined with FasterRCNN model as claimed in claim 7, characterized in that: Determining the pixel area at the initial moment based on the pixel corresponding to the state prediction point at the initial moment includes: The pixel point area at the initial moment is generated with the pixel point corresponding to the state prediction point at the initial moment as the center.

9. A track initiation method based on a combination of rules and FasterRCNN model according to any one of claims 6 to 8, characterized in that: When training the FasterRCNN: In the radar echo image used as a training sample, the track start labeling box is larger than the track trajectory.

10. A track initiation device based on a combination of rules and a FasterRCNN model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the track initiation method based on the combination of rules and FasterRCNN model described in any one of claims 1 to 9 is implemented.

Images