Radar track-before-detection device and method based on semantic prompt network

Abstract

The invention provides a radar track-before-detect device and method based on a semantic prompt network. The device comprises an efficient universal segmentation model, a feature pool, a peak convolutional network model and a focusing module. The efficient universal segmentation model comprises a prompt encoder, a data encoder and a prompt guide decoder; the prompt encoder is used for obtaining prompt encoding features and outputting the prompt encoding features to the prompt guide decoder; the data encoder is used for obtaining data encoding features; under the condition that the target is detected, the target is output to the prompt guiding decoder; the prompt guide decoder is used for outputting the target candidate area at the current moment to the feature pool and outputting the relative position of the target candidate area at the current moment relative to the observed position at the current moment to the peak convolutional network model; the feature pool is used for outputting M frames of target candidate region batch data to the peak convolutional network model; the peak convolutional network model is used for acquiring a target state finally estimated at the current moment and outputting the target state to the focusing module; and the focusing module is used for acquiring prompt information at the next moment.

Description

Technical Field

[0001] The present invention relates to the technical field of target tracking before detection, and in particular to a radar tracking before detection device and method based on a semantic prompt network. Background Art

[0002] Traditional detection and tracking algorithms are widely used in radar, sonar, and remote sensing. In radar systems, methods based on the generalized likelihood ratio test (GLRT) only retain binary results, discarding the target's kinematic, radiometric, and micro-Doppler information. Consequently, these algorithms have limitations at low signal-to-noise ratios (SNRs).

[0003] Tracking Before Detection (TBD) is an effective method to improve the detectability of weak targets. In TBD methods, raw or minimally processed sensor data is passed directly to the tracking stage to minimize information loss in the detection stage. State-of-the-art TBD methods can be divided into batch methods and sequential Bayesian estimation methods. On the one hand, sequential Bayesian estimation methods include random vector-based methods and random finite set (Random Field Set, RFS)-based methods. On the other hand, batch methods estimate the target state through Maximum Likelihood Probabilistic Data Association (ML-PDA), Hough Transform (Hough Transform) or Dynamic Programming (DP). These methods can improve detection performance by exploiting the spatiotemporal correlation between multiple frames and integrating the target energy (or maximum a posteriori probability).

[0004] There have been many studies on detection and tracking, including popular research topics such as nonlinear tracking, heterogeneous clutter, multi-target tracking, threshold setting, maneuvering targets, and short trajectory filtering. However, the main disadvantages of detection and tracking technology are its high implementation cost and computational complexity. In radar systems, each frame of data is a high-dimensional data matrix or tensor, which means that the number of parameters can easily reach millions during joint batch processing. For example, radar detection problems require processing four-dimensional tensors (range, Doppler, azimuth, and elevation) or even higher-dimensional tensors. In order to obtain accurate estimation results, more than a thousand resolution units are usually required in the range and azimuth dimensions. This can lead to huge demands on computational and memory resources.

[0005] To improve the efficiency of multi-frame tracking before detection (MF-TBD), a two-stage tracking-before-detection architecture is one of the most commonly discussed approaches. In this approach, the preprocessing step uses a traditional constant false alarm rate (CFAR) algorithm with a relatively low threshold and a pattern extractor to generate candidate detections. Only measurements with high intensity are processed further, which helps reduce computational complexity. However, a major drawback of these methods is that they remain extremely dependent on target energy. Under low signal-to-noise ratio conditions, many false alarms will be included in the calculation, resulting in degraded performance. Summary of the Invention

[0006] The present invention provides a radar pre-detection tracking device and method based on a semantic cue network, which can use the tracking results of each frame to estimate the target appearance position in the next frame, provide prior information for detection, and further improve the accuracy of detection and tracking.

[0007] According to one aspect of the present invention, a radar tracking before detection device based on a semantic cue network is provided, the device comprising an efficient general segmentation model, a feature pool, a spiking convolutional network model and a focusing module;

[0008] The efficient general segmentation model includes a cue encoder, a data encoder, and a cue-guided decoder;

[0009] The prompt encoder is used to receive prompt information at a current moment, encode the prompt information at the current moment, obtain prompt coding features, and output the features to the prompt guidance decoder; wherein the prompt information at the current moment includes tracking information at a previous moment;

[0010] The data encoder is used to receive an observation at a current moment and encode the observation at the current moment to obtain a data encoding feature; it is also used to determine whether the target is detected based on the M frames of target state batch data and the observation batch data, and if the target is detected, output the data encoding feature to the prompt guidance decoder; wherein the observation is range-Doppler data, the M frames of target state batch data include the target state at the current moment and the target state at M-1 consecutive moments before the current moment; the M frames of observation batch data include the observation at the current moment and the observation at M-1 consecutive moments before the current moment;

[0011] The cue-guided decoder is used to decode the cue encoding feature and the data encoding feature to obtain the target candidate area at the current moment and the relative position of the target candidate area at the current moment relative to the observation at the current moment, and output the target candidate area at the current moment to the feature pool, and output the relative position of the target candidate area at the current moment relative to the observation at the current moment to the spike convolutional network model;

[0012] The feature pool is used to store the target candidate area at each moment, and output M frames of target candidate area batch data to the spike convolutional network model; wherein the M frames of batch data include the target candidate area at the current moment and the target candidate areas at M-1 consecutive moments before the current moment;

[0013] The spike convolutional network model is used to obtain a preliminary estimated target state at the current moment based on M frames of batch data; and is also used to obtain a final estimated target state at the current moment based on the relative position of the target candidate area at the current moment relative to the observation at the current moment and the preliminary estimated target state at the current moment, and output it to the focusing module;

[0014] The focusing module is used to obtain prompt information at the next moment based on the final estimated target state at the current moment and the observation at the next moment.

[0015] Preferably, the data encoder includes a subgraph partitioning module, a linear embedding coding module, a position coding module, a first-layer visual attention module, a second-layer visual attention module, a third-layer visual attention module, and a fourth-layer visual attention module, which are connected in sequence; the four-layer visual attention modules have the same structure.

[0016] Preferably, the prompt-guided decoder includes a first layer of dual-path attention module, a second layer of dual-path attention module, and a convolutional layer connected in sequence; the two layers of dual-path attention modules have the same structure.

[0017] Preferably, the target state preliminarily estimated at the current moment is obtained by the following formula:

[0018]

[0019] Where, represents the target state initially estimated at time k, PKC(·) represents the spike convolutional network model, S k-M+1:k represents the target candidate area from k-M+1 to time k, Θ k represents the temporal state space of all target motions from k-M+1 to time k, represents the target state initially estimated from k-M+1 to time k, Indicates that along k-M+1 to k moment, based on S k-M+1:k Under the condition of x k-M+1:k The metric integral of x k-M+1:k represents the target state from k-M+1 to time k, and argmax(·) represents the value obtained when “·” is maximum.

[0020] Preferably, the prompt information at the next moment is obtained by the following formula:

[0021]

[0022] Where V k+1 represents the prompt information at time k+1, Fo(·) represents the focus module, x k represents the target state at time k, p(Z k+1 |x k ) represents the target state x at time k k Converted to observation Z at time k+1 k+1 The probability of Z k+1 represents the observation at time k+1.

[0023] Preferably, whether the target is detected is determined by the following formula:

[0024]

[0025] Where, Indicates that along k-M+1 to k moment, based on Z k-M+1:k Under the condition of x k-M+1:k The metric integral, Z k-M+1:k represents the observation from time k-M+1 to time k, η represents the preset threshold, D1 represents the case where the target is detected, and D0 represents the case where the target is not detected.

[0026] Preferably, the prompt information is a "click prompt" or a "box prompt".

[0027] According to another aspect of the present invention, a radar tracking-before-detection method based on a semantic cue network is provided. The method uses any of the above-mentioned devices to perform tracking-before-detection, and the method includes:

[0028] The prompt encoder of the efficient general segmentation model receives the prompt information at the current moment, encodes the prompt information at the current moment, obtains the prompt encoding feature and outputs it to the prompt-guided decoder of the efficient general segmentation model;

[0029] The data encoder of the efficient general segmentation model receives the observation at the current moment and encodes the observation at the current moment to obtain the data encoding feature. It also determines whether the target is detected based on the M-frame target state batch data and the observation batch data. If the target is detected, it outputs the data encoding feature to the cue-guided decoder.

[0030] The cue-guided decoder decodes the cue encoding features and the data encoding features to obtain the target candidate at the current moment and the relative position of the target candidate at the current moment relative to the observation at the current moment, and outputs the target candidate at the current moment to the feature pool, and outputs the relative position of the target candidate at the current moment relative to the observation at the current moment to the spiking convolutional network model;

[0031] The feature pool stores the target candidate regions at each moment and outputs the M-frame target candidate region batch data to the spiking convolutional network model;

[0032] The spiking convolutional network model obtains a preliminary estimate of the target state at the current moment based on the M-frame batch data. It also obtains the final estimate of the target state at the current moment based on the relative position of the target candidate area at the current moment relative to the current moment observation and the preliminary estimate of the target state at the current moment and outputs it to the focusing module.

[0033] The focusing module obtains the prompt information at the next moment based on the final estimated target state at the current moment and the observation at the next moment.

[0034] According to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above methods when executing the computer program.

[0035] The technical solution of the present invention is applied to improve the efficient general segmentation model and, after compromise, reduce the patch dimension in the encoder to 4 from the perspectives of performance and efficiency to improve efficiency and accuracy. At the same time, different types of point and box prompts are generated to enhance the network's generalization ability and cope with different signal-to-noise ratio conditions and tracking capabilities in complex situations. In addition, the target motion information and state change information between multiple frames are fully considered, and the dimensional convolution module in the spike convolution network model is used to capture related information from aspects such as target energy and target motion state, thereby improving the estimation accuracy of the target state. At the same time, the present invention provides a fine-tuned visual large model and an efficient general segmentation model, and comprehensively considers processing efficiency and accuracy to build a model architecture that is most suitable for pre-detection tracking tasks. Furthermore, the present invention combines the spike convolution network model with the efficient general segmentation model into a new network, which can not only quickly process radar range-Doppler data, but also fully utilize the prior information returned from the previous tracking stage to accurately estimate the target state in multiple frames. The present invention efficiently utilizes a neural network to achieve end-to-end detection and tracking of radar targets, processes tracking information in a "prompt" manner, and generates a priori conditions for detection at the next moment. The present invention can be directly applied to Doppler radar equipment, such as continuous wave, pulse Doppler, and other radar equipment with various detection systems, including a signal processing module running on a radar processor or a computer program encoded on a computer storage medium. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings are included to provide a further understanding of the embodiments of the present invention, constitute a part of the specification, illustrate the embodiments of the present invention, and together with the description, explain the principles of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0037] Figure 1 A schematic structural diagram of a radar tracking before detection device based on a semantic prompt network provided according to an embodiment of the present invention is shown;

[0038] Figure 2 Shown Figure 1 Schematic diagram of the structure of the efficient general segmentation model in;

[0039] Figure 3 Shown Figure 1 The working principle diagram of the focusing module in ;

[0040] Figure 4 A workflow diagram of a radar tracking-before-detection method based on a semantic hint network is shown according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is in no way intended to limit the present invention and its application or use. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0042] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0043] Unless otherwise specifically stated, the relative arrangement of the parts and steps, the numerical expressions and the numerical values ​​set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The techniques, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods and equipment should be considered as part of the authorization specification. In all examples shown and discussed here, any specific values ​​should be interpreted as being merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following figures, and therefore, once an item is defined in one figure, it does not need to be further discussed in subsequent figures.

[0044] Based on the principle of compressed sensing, the low-rank characteristics of the echo data matrix and the sparsity of the target can be utilized to improve efficiency. The target unit of interest only occupies a small part of the tens of thousands of range resolution units. Direct batch processing of raw data may result in a large amount of resource consumption. Therefore, the present invention proposes a pre-detection tracking method based on a semantic cue network. Before batch processing, the data will first undergo local feature selection. In this step, key information of the target, such as motion, distance, and Doppler shift, will be retained. This method reduces the requirements for computing and memory resources while maintaining pre-detection tracking accuracy.

[0045] The tracking-before-detection method proposed in this invention consists of two key components: coarse detection and fine tracking. The coarse detection component predicts the target's likely location in the next time step based on the current observation data. In contrast, the fine tracking component leverages the output of the coarse detection from the previous time step to rapidly estimate the target's multi-frame state (i.e., its trajectory). To achieve their intended functions, these two components must possess specific capabilities. The coarse detection component must rapidly process a large amount of observation data while effectively integrating feedback from the fine tracking component. In this context, the structure and functional characteristics of the Efficient Segment Anything Model (SAM) are well suited to meet the requirements of coarse detection. The SAM is equipped with two input interfaces: a data interface for receiving observation data and a hint interface for interpreting the tracking results from the previous time step. However, the SAM was designed specifically for image segmentation tasks. Its advantages are evident after training on a large dataset containing 11 million images and 1 billion masks. However, due to its large parameter size and high computational resource consumption, the SAM's applicability is very limited. Therefore, this invention improves upon the SAM. On the other hand, the precise tracking component focuses on analyzing the target's features and relationship information across multiple frames to accurately estimate its trajectory. This component uses a Peak ConvNetwork (PKC) model specifically designed to process Doppler range matrices, enabling it to fully capture the semantic information in radar data.

[0046] In order to have a further understanding of the present invention, the following Figure 1-Figure 4 The radar pre-detection tracking device and method based on semantic prompt network of the present invention are described in detail.

[0047] Assuming that the target moves in a d-dimensional space, the observation obtained at the kth moment is The observation value of the jth resolution unit in the observation is z k,j , then the k-time observation can be expressed as:

[0048] Z k ={z k,j ,j=1,2,...,N}

[0049] Where N=N1·N2·...·N d ,N1,N2,...,N d Represents the obtained observation Z k The number of resolution units in the d-dimensional state space. Assuming that the target occupies only one distance unit in the narrowband detection and tracking process, for binary hypothesis testing, if the target exists in the observation, it can be expressed as follows:

[0050]

[0051] Among them, p(Z k |x k ) is the target state x at time k k Converted to observation Z k The probability, p(z k,j |x k ) is the probability that the target state is transferred to the jth resolution unit in the observation. If the target is not in the jth resolution unit, p(z k,j |H0) indicates that p(z k,j |H0) represents the probability that the target is not in the jth resolution unit at time k. If the target does not exist, it is expressed as follows:

[0052]

[0053] where p(Z k |H0) represents the probability that the target does not exist at time k, that is, there is no target in any of the resolution units 1 to N.

[0054] To describe the target motion process, we use the target state transfer equation and the measurement equation. For a linear process, the former can be written as:

[0055] x k =Fx k-1 +ω k

[0056] where ω k is the system processing noise at time k, and F is the state transfer function, which can be expressed as follows in the linear Kalman filter process:

[0057]

[0058] Among them I d is the identity matrix, T is the observation time interval, is the Kronecker product. The measurement equation can be expressed as:

[0059] Z k =Hx k +μ k

[0060] where μ k is the observation noise, H is the observation function, and in the first-order linear Kalman filter process, it can be expressed as follows:

[0061]

[0062] For the multi-frame target tracking process, it can be regarded as an estimation of the target state of consecutive frames, and the result of parameter estimation is the target track. We process M frames of data together and regard them as a batch of data, which is recorded as Z k-M+1:k , then the detection process of this timing signal can be written as follows:

[0063] where Θ k represents the temporal state space of all target motions from time k - M + 1 to time k. η is a preset threshold, D1 represents the case where the target is detected, and D0 represents the case where the target is not detected. In the case of D1, the proposed algorithm outputs target track information; otherwise, it is assumed that there is no target and no track. It is along k-M+1 to time k, based on observation Z k-M+1:k Under the condition of x k-M+1:k The metric integral of x k-M+1:k is the probability of the state sequence from the target. We traverse all possibilities of this metric integral and take the maximum value to compare with the preset threshold η. If it is greater than the threshold, it is considered that the target exists in the multi-frame observation, otherwise the target does not exist. This metric can be expressed in the joint probability distribution (JPD) as:

[0064]

[0065] when When the target state is maximized, it is the target state estimate under the maximum a posteriori criterion. At this time, the estimate of the target continuous state can be described as the value of the target state when the metric function is maximized. However, in practical applications, due to the complex scattering of the target and the clutter environment, the posterior probability is often very complex and difficult to estimate. Therefore, this paper provides an estimation method using a semantically hinted neural network.

[0066] like Figure 1 As shown, the present invention provides a radar pre-detection tracking device based on a semantic cue network, the device comprising an efficient general segmentation model, a feature pool, a spike convolutional network model and a focusing module;

[0067] The efficient general segmentation model includes a cue encoder, a data encoder, and a cue-guided decoder;

[0068] The prompt encoder is used to receive prompt information at a current moment, encode the prompt information at the current moment, obtain prompt coding features, and output the features to the prompt guidance decoder; wherein the prompt information at the current moment includes tracking information at a previous moment;

[0069] The data encoder is used to receive an observation at a current moment and encode the observation at the current moment to obtain a data encoding feature; it is also used to determine whether the target is detected based on the M frames of target state batch data and the observation batch data, and if the target is detected, output the data encoding feature to the prompt guidance decoder; wherein the observation is range-Doppler data, the M frames of target state batch data include the target state at the current moment and the target state at M-1 consecutive moments before the current moment; the M frames of observation batch data include the observation at the current moment and the observation at M-1 consecutive moments before the current moment;

[0070] The cue-guided decoder is used to decode the cue encoding feature and the data encoding feature to obtain the target candidate area at the current moment and the relative position of the target candidate area at the current moment relative to the observation at the current moment, and output the target candidate area at the current moment to the feature pool, and output the relative position of the target candidate area at the current moment relative to the observation at the current moment to the spike convolutional network model;

[0071] The feature pool is used to store the target candidate area at each moment, and output M frames of target candidate area batch data to the spike convolutional network model; wherein the M frames of batch data include the target candidate area at the current moment and the target candidate areas at M-1 consecutive moments before the current moment;

[0072] The spike convolutional network model is used to obtain a preliminary estimated target state at the current moment based on M frames of batch data; and is also used to obtain a final estimated target state at the current moment based on the relative position of the target candidate area at the current moment relative to the observation at the current moment and the preliminary estimated target state at the current moment, and output it to the focusing module;

[0073] The focusing module is used to obtain prompt information at the next moment based on the final estimated target state at the current moment and the observation at the next moment.

[0074] According to one embodiment of the present invention, Figure 2 The improved efficient general segmentation model is shown in Figure 2. The efficient general segmentation model has two input layers, namely the prompt encoder input layer and the data encoder input layer, which are used to process the prompt information V k and range-Doppler data Z k , and get the hint code and data code. Their role is to convert the input data into a code that the network can understand and preliminarily extract the range-Doppler data Z kThe hint encoder is mainly composed of linear embedding coding, and the data encoder is mainly composed of sub-graph partitioning module, linear embedding coding module, position coding module, and 4-layer visual attention module. The hint coding features and data coding features output by them are sent to the hint-guided decoder together. The hint-guided decoder is composed of 2 layers of dual-path attention modules and 1 layer of 2D convolution layer. It is a lightweight decoder. The decoder will output the target candidate area S at the current moment k and record target candidate area S k Relative observation Z k Position O k , so the efficient general segmentation model completes the process from raw range-Doppler data to feature extraction, which can be expressed as follows:

[0075] {S k ,O k = SAM(Z k ,V k )

[0076] Among them O k Is the predicted feature S k Relative observation Z k The relative position of S is used to convert the subsequent estimation results into the global estimation. k We take the region with higher prediction probability as the target candidate region, that is, the feature. k is the prompt information at the kth moment. The network can accept two prompt modes: "point prompt" and "box prompt". They contain the tracking information of the previous moment. The model encodes and understands it and constructs the prior distribution of the target state at that moment. V k The tracking result of the previous moment is obtained after "focusing". The working principle of the focusing module is as follows: Figure 3 As shown, its function is to place the result of the prediction at the previous moment in the center of the prompt at the next moment, playing a "focusing" role. The process can be written as:

[0077]

[0078] Where V k+1 represents the prompt information at time k+1, Fo(·) represents the focus module, x k represents the target state at time k, p(Z k+1 |x k ) represents the target state x at time k k Converted to observation Z at time k+1 k+1 The probability of Z k+1 represents the observation at time k+1.

[0079] Among them, the focusing module predicts the target observation position at the next moment under given current state conditions and takes the value when the probability is maximum. If the target motion pattern obeys the Poisson distribution, the module degenerates into an elliptical wave gate in the Kalman filtering process.

[0080] The spiking convolutional network model is used to process multi-frame data. This method outperforms most current neural networks in understanding and processing radar range-Doppler data, and is therefore selected for accurate target state estimation. Previously, this method was only used for detection, but this invention applies it to tracking. The process by which the network estimates the target state can be described as follows:

[0081]

[0082] Where, represents the target state initially estimated at time k, PKC(·) represents the spike convolutional network model, S k-M+1:k represents the target candidate area from k-M+1 to time k, Θ k represents the temporal state space of all target motions from k-M+1 to time k, represents the target state initially estimated from k-M+1 to time k, Indicates that along k-M+1 to k moment, based on S k-M+1:k Under the condition of x k-M+1:k The metric integral of x k-M+1:k represents the target state from k-M+1 to time k, and argmax(·) represents the value obtained when “·” is maximum.

[0083] Combining the relative position information with the estimation results, the final estimated state can be obtained as follows:

[0084]

[0085] in The initial target state estimated by the spike convolution module is represented by the position of the target in the candidate area. To obtain the final target state in the complete observation data, the relative position information O needs to be added. k , we describe the process as above.

[0086] The efficient general segmentation model and the spike convolutional network model of the present invention respectively replace the preprocessing step and the multi-frame detector in the two-step detection-before-tracking method. Figure 4 shown.

[0087] In the coarse detection component, we use a lightweight and efficient universal segmentation model to replace traditional low-threshold preprocessing methods. This approach helps efficiently analyze target features and accurately determine the target area. In the fine tracking component, we use a spiking convolutional model to replace traditional algorithms such as dynamic programming (DP). This method can analyze target association information across multiple frames and provide accurate target trajectory estimation.

[0088] In summary, the present invention provides a radar pre-detection tracking device and method based on a semantic cueing network. The device and method consist of an improved efficient universal segmentation model and a PeakConv-based radar semantic segmentation network (PKC network). The efficient universal segmentation model predicts the target state at the next moment based on the current input, while the PKC model uses the efficient universal segmentation model at the previous moment for precise tracking. This achieves efficient, computationally efficient, and highly accurate low-SNR target detection and tracking. This approach has the following beneficial effects:

[0089] 1. A higher detection probability can be achieved. The proposed method extracts more potential information, thereby improving the detection probability;

[0090] 2. More accurate state estimation. The proposed method uses a 3D convolution module to capture the changes in target state between consecutive frames, rather than just focusing on energy accumulation, thereby achieving more accurate estimation.

[0091] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used herein to describe the spatial positional relationship of a device or feature to other devices or features as shown in the figures. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figures. For example, if the device in the drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.

[0092] In addition, it should be noted that the use of terms such as "first" and "second" to limit components is only for the convenience of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of the present invention.

[0093] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A radar tracking device based on semantic hint network, characterized in that: The device includes an efficient general segmentation model, a feature pool, a spiking convolutional network model and a focusing module; The efficient general segmentation model includes a cue encoder, a data encoder, and a cue-guided decoder; The prompt encoder is used to receive prompt information at a current moment, encode the prompt information at the current moment, obtain prompt coding features, and output the features to the prompt guidance decoder; wherein the prompt information at the current moment includes tracking information at a previous moment; The data encoder is used to receive an observation at a current moment and encode the observation at the current moment to obtain a data encoding feature; it is also used to determine whether the target is detected based on the M frames of target state batch data and the observation batch data, and if the target is detected, output the data encoding feature to the prompt guidance decoder; wherein the observation is range-Doppler data, the M frames of target state batch data include the target state at the current moment and the target state at M-1 consecutive moments before the current moment; the M frames of observation batch data include the observation at the current moment and the observation at M-1 consecutive moments before the current moment; The cue-guided decoder is used to decode the cue encoding feature and the data encoding feature to obtain the target candidate area at the current moment and the relative position of the target candidate area at the current moment relative to the observation at the current moment, and output the target candidate area at the current moment to the feature pool, and output the relative position of the target candidate area at the current moment relative to the observation at the current moment to the spike convolutional network model; The feature pool is used to store the target candidate area at each moment, and output M frames of target candidate area batch data to the spike convolutional network model; wherein the M frames of batch data include the target candidate area at the current moment and the target candidate areas at M-1 consecutive moments before the current moment; The spike convolutional network model is used to obtain a preliminary estimated target state at the current moment based on M frames of batch data; and is also used to obtain a final estimated target state at the current moment based on the relative position of the target candidate area at the current moment relative to the observation at the current moment and the preliminary estimated target state at the current moment, and output it to the focusing module; The focusing module is used to obtain prompt information at the next moment based on the final estimated target state at the current moment and the observation at the next moment.

2. The device according to claim 1, characterized in that The data encoder includes a subgraph partitioning module, a linear embedding coding module, a position coding module, a first-layer visual attention module, a second-layer visual attention module, a third-layer visual attention module, and a fourth-layer visual attention module connected in sequence; the four-layer visual attention modules have the same structure.

3. The device according to claim 1, characterized in that The prompt-guided decoder includes a first-layer dual-path attention module, a second-layer dual-path attention module, and a convolutional layer connected in sequence; the two layers of dual-path attention modules have the same structure.

4. The device according to claim 1, characterized in that The target state of the current moment is obtained by the following formula: Where, represents the target state initially estimated at time k, PKC(·) represents the spike convolutional network model, S k-M+1:k represents the target candidate area from k-M+1 to time k, Θ k represents the temporal state space of all target motions from k-M+1 to time k, represents the target state initially estimated from k-M+1 to time k, Indicates that along k-M+1 to k moment, based on S k-M+1:k Under the condition of x k-M+1:k The metric integral of x k-M+1:k represents the target state from k-M+1 to time k, and argmax(·) represents the value obtained when "·" is maximum.

5. The device according to claim 1, characterized in that Get the prompt information for the next moment through the following formula: Where V k+1 represents the prompt information at time k+1, Fo(·) represents the focus module, x k represents the target state at time k, p(Z k+1 |x k ) represents the target state x at time k k Converted to observation Z at time k+1 k+1 The probability of Z k+1 represents the observation at time k+1.

6. The device according to claim 1, characterized in that Use the following formula to determine whether the target is detected: Where, Indicates that along k-M+1 to k moment, based on Z k-M+1:k Under the condition of x k-M+1:k The metric integral, Z k-M+1:k represents the observation from time k-M+1 to time k, η represents the preset threshold, D1 represents the case where the target is detected, and D0 represents the case where the target is not detected.

7. The device according to claim 1, characterized in that The prompt information is a "point prompt" or a "box prompt".

8. A radar tracking before detection method based on semantic cue network, characterized in that: The method uses the device according to any one of claims 1 to 7 to perform pre-detection tracking, and the method comprises: The prompt encoder of the efficient general segmentation model receives the prompt information at the current moment, encodes the prompt information at the current moment, obtains the prompt encoding feature and outputs it to the prompt-guided decoder of the efficient general segmentation model; The data encoder of the efficient general segmentation model receives the observation at the current moment and encodes the observation at the current moment to obtain the data encoding feature. It also determines whether the target is detected based on the M-frame target state batch data and the observation batch data. If the target is detected, it outputs the data encoding feature to the cue-guided decoder. The cue-guided decoder decodes the cue encoding features and the data encoding features to obtain the target candidate at the current moment and the relative position of the target candidate at the current moment relative to the observation at the current moment, and outputs the target candidate at the current moment to the feature pool, and outputs the relative position of the target candidate at the current moment relative to the observation at the current moment to the spiking convolutional network model; The feature pool stores the target candidate regions at each moment and outputs the M-frame target candidate region batch data to the spiking convolutional network model; The spiking convolutional network model obtains a preliminary estimate of the target state at the current moment based on the M-frame batch data. It also obtains the final estimate of the target state at the current moment based on the relative position of the target candidate area at the current moment relative to the current moment observation and the preliminary estimate of the target state at the current moment and outputs it to the focusing module. The focusing module obtains the prompt information at the next moment based on the final estimated target state at the current moment and the observation at the next moment.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images