Sea surface small target detection and tracking method and system based on multi-frame detection

Abstract

The invention provides a sea surface small target detection and tracking method and system based on multi-frame detection, and the method proposes an overall detection network MF-IFA-UNet, the network is based on a UNet encoder-decoder architecture, and an inter-frame attention mechanism module IFA is embedded in an encoder, so that the network has the capability of paying attention to target features at different moments, and the detection accuracy is improved. Meanwhile, a small area detection model SADN is provided and is used for carrying out secondary search on the position where the target possibly appears. When target detection is carried out on an RD image to be detected, a target track is obtained through MF-IFA-UNet matching, region segmentation is carried out on a lost target, an obtained local region small-size RD image is input into SADN, through the two-stage method, even if the target is not detected in a certain frame due to clutter interference, the target still has an opportunity to be found again in a local range, and the detection accuracy is improved. The track interruption is effectively avoided, and the stability and precision of multi-target tracking are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target detection, and in particular to a method and system for detecting and tracking small targets on the sea surface based on multi-frame detection. Background Art

[0002] Detecting and tracking small, maneuvering targets on the sea surface is a crucial function of ocean surveillance radars. However, traditional methods have significant limitations in high sea conditions and complex sea clutter. Small targets on the sea surface are characterized by their small size, variable trajectory, and low signal-to-clutter ratio (SCR). Radar echoes are easily overwhelmed by clutter such as waves, making them difficult to detect and track in a timely and reliable manner. The commonly used "detect-before-track" (DBT) method in existing technologies detects targets independently in each radar image frame and then associates their tracks. However, because it fails to fully utilize inter-frame correlation information, the DBT method is prone to generating a large number of false tracks in low SCR and high clutter environments, resulting in a high miss detection rate and poor tracking stability. Another "track-before-detect" (TBD) method attempts to simultaneously detect and track small targets by accumulating energy across multiple frames before threshold detection. This method directly estimates the target presence probability based on multiple frames of data, improving the detection probability of small, weak targets. For example, the particle filter TBD method (PF-TBD) and the TBD method based on multidimensional Hough transform have improved the weak target detection performance to a certain extent; however, PF-TBD suffers from serious particle degradation problems and limited tracking accuracy in strong clutter; Hough transform TBD has high computational complexity and is sensitive to noise, and dynamic programming TBD is prone to produce false trajectories when the target signal is extremely weak.

[0003] In recent years, the application of deep learning to small target detection has become a trend, such as using convolutional neural networks to extract features and improve detection robustness. However, existing deep learning methods often only utilize single frames or simply accumulate information, ignoring the temporal and spatial correlation between targets and clutter. As a result, detection performance in non-stationary sea clutter remains unsatisfactory. Summary of the Invention

[0004] Purpose of the invention: The first purpose of the present invention is to provide a method for detecting and tracking small targets on the sea surface based on multi-frame detection, so as to realize continuous detection and tracking of small targets on the sea surface; the second purpose is to provide a system for detecting and tracking small targets on the sea surface based on multi-frame detection.

[0005] Technical solution: A method for detecting and tracking small targets on the sea surface based on multi-frame detection, comprising the following steps:

[0006] S1. Add a simulated target to the sea clutter background, then cut the data matrix after adding the simulated target into a set size and convert it to obtain a simulated RD map, and cut the area containing the simulated target in the simulated RD map into a set size to obtain a local RD map;

[0007] S2, preprocessing the simulated RD map and the local RD map respectively;

[0008] S3. Design the overall detection network MF-IFA-UNet, input the preprocessed simulated RD map into MF-IFA-UNet for training and save the weights; design the small area detection network SADN, input the preprocessed local RD map into SADN for training and save the weights;

[0009] S4. Input the RD maps to be detected accumulated from several frames into MF-IFA-UNet, extract the spatiotemporal features between frames, and output a detection probability heat map. Cluster the values judged as true targets in the detection probability heat map to obtain the true target center position in each frame, and output the detection result of each true target. If the true target appears for the first time, the current tracking sequence is empty, and a new tracking sequence is initialized for each detection result.

[0010] S5. Based on the center position of the real target output by MF-IFA-UNet, the Kalman filter is used to predict the position of the real target in the current frame based on the position and velocity of the real target in the previous frame. The current frame detection result is then correlated and matched with the predicted position. If the match is successful, the target trajectory is updated. If the match fails, the predicted position is used as the center to crop the local RD map to be detected.

[0011] S6. Input the local RD map to be detected obtained in S5 into the SADN, output the detection probability heat map, perform a clustering operation to determine whether there is a real target that failed to be matched in S5 in the local RD map to be detected. If the real target is detected, the real target is re-included in the tracking sequence and the target trajectory is updated;

[0012] S7. Repeat steps S4 to S6 to detect and track the real target frame by frame.

[0013] Specifically, in step S1, the model formula of sea clutter and simulated target is:

[0014]

[0015] Where: H0 is the model containing only sea clutter, H1 is the model containing simulated target echo and sea clutter, z(m,p) is the complex time series received by the radar at cell m, z p(m,p) is the sea clutter time series at the reference cells around cell m, A0(m,p) is the complex coefficient amplitude matrix, j is the imaginary unit, R(p) is the radial range of the target at the pth pulse, and λ is the radar wavelength;

[0016] The dynamic formula for simulating target motion is:

[0017]

[0018] Where: r (p) is the instantaneous radial velocity, a r (p) is the instantaneous acceleration, a r (p) obeys the normal distribution with mean μ r , with a standard deviation of δ r ; Δt is the sampling time.

[0019] Specifically, in step S3, the overall detection network MF-IFA-UNet includes an encoder and a decoder. The encoder includes several convolutional layers and downsampling layers, and uses three-dimensional convolution to extract the joint features of the spatial and temporal dimensions of the input multi-frame RD graph; the decoder includes a multi-level transposed convolutional layer symmetrical to the encoder, and fuses the features of the corresponding convolutional layer of the encoder with the decoding features of the transposed convolutional layer through jump connections, and outputs a detection probability heat map with the same size as the input RD graph. The overall detection network MF-IFA-UNet embeds the inter-frame attention mechanism module IFA during the convolution process.

[0020] Specifically, in step S3, the small area detection network SADN includes a downsampling encoder and an upsampling decoder. The downsampling encoder includes several levels of residual downsampling modules, each level of residual downsampling module includes two convolutional layers and a residual connection; the upsampling decoder includes a multi-level transposed convolutional layer symmetrical to the downsampling encoder, and fuses the features of the corresponding convolutional layer of the downsampling encoder with the decoded features of the transposed convolutional layer through jump connections, and outputs a detection probability heat map with the same size as the input simulated RD map.

[0021] Specifically, in the downsampling encoder, the coordinate attention CA and efficient channel attention ECA mechanism modules are introduced after each level of downsampling residual downsampling module.

[0022] Specifically, in step S5, the Hungarian algorithm is used to associate and match the current frame detection result with the predicted position:

[0023] The Mahalanobis distance between the center position of the true target in the detection result and the corresponding predicted position of the true target is calculated as the value of the cost matrix and assigned using the Hungarian algorithm; the true target whose Mahalanobis distance with the predicted position in the current frame is within the preset threshold is assigned to the corresponding track, and the target matching is determined to be successful; otherwise, the true target is determined to be lost.

[0024] Specifically, step S6 includes:

[0025] A clustering operation is performed on the detection probability heat map to obtain the center position of the true target. When the true target is not empty, the Mahalanobis distance between the true target center position and the predicted position is calculated. When the Mahalanobis distance is less than the preset threshold, the target is judged to be matched successfully. Otherwise, the match fails and the frame is recorded as a lost frame. When the number of consecutive lost frames exceeds the set threshold, it is judged that the true target trajectory has disappeared.

[0026] Specifically, in step S1, the generation process of the simulated RD graph is as follows: performing a sliding window operation on the data matrix after adding the simulated target, and then performing a fast Fourier transform and shifting the frequency center to generate the simulated RD graph.

[0027] Specifically, in step S2, the preprocessing includes: performing channel compression on the simulated RD map and the local RD map, converting them into one-dimensional grayscale images, and normalizing the pixels in the simulated RD map and the local RD map.

[0028] The present invention also provides a sea surface small target detection and tracking system based on multi-frame detection, comprising:

[0029] RD map generation module: used to add simulated targets in the sea clutter background, then cut the data matrix after adding the simulated targets into a set size and convert it into a simulated RD map, and cut the area containing the simulated targets in the simulated RD map into a set size to obtain a local RD map;

[0030] Preprocessing module: used to preprocess RD graph and local RD graph respectively;

[0031] Network training module: used to design the overall detection network MF-IFA-UNet, input the preprocessed simulated RD map into MF-IFA-UNet for training and save the weights; design the small area detection network SADN, input the preprocessed local RD map into SADN for training and save the weights;

[0032] Target localization module: This module is used to input the RD maps to be detected accumulated over several frames into MF-IFA-UNet, extract the spatiotemporal features between frames, and output a detection probability heat map. It then performs a clustering operation on the values judged to be true targets in the detection probability heat map to obtain the true target center position for each frame and output the detection result for each true target. If a true target appears for the first time, the current tracking sequence is empty, and a new tracking sequence is initialized for each detection result.

[0033] Target matching module: It is used to predict the position of the real target in the current frame based on the center position of the real target output by MF-IFA-UNet and the position and velocity of the real target in the previous frame using the Kalman filter. The detection result of the current frame is then matched with the predicted position. If the match is successful, the target trajectory is updated. If the match fails, the local RD map to be detected is cropped with the predicted position as the center.

[0034] Lost Target Matching Module: This module is used to input the local RD map to be detected obtained by the target matching module into the SADN, output a detection probability heat map, and perform a clustering operation to determine whether there is a real target in the local RD map to be detected that failed to be matched in the target matching module. If the real target is detected, it is re-included in the tracking sequence and the target trajectory is updated;

[0035] Continuous detection and tracking module: used to repeatedly execute the target positioning module, target matching module and lost target matching module in sequence, and detect and track the target frame by frame.

[0036] Beneficial effect: Compared with the prior art, the significant effect of the present invention is that the present invention obtains a large-scale RD map and a small-scale RD map containing the simulated target by adding a simulated target to a pure sea clutter background, and designs detection networks for these two types of targets to be detected respectively. For the large-scale RD map, the present invention proposes an overall detection network MF-IFA-UNet, which is based on the UNet encoder-decoder architecture, and embeds an inter-frame attention mechanism module IFA in the encoder, so that the network has the ability to pay attention to target features at different times. When the target is clear in the current frame but partially submerged by clutter in the adjacent frame, the network can still capture the target based on the inter-frame correlation. This mechanism effectively reduces the missed detection rate and false detection rate of multi-frame detection. For the small-scale RD map, the present invention proposes a small area detection model SADN for secondary search of the location where the target may appear. SADN can realize accurate detection of weak targets in the small-scale RD map. When the target disappears in the global detection due to clutter occlusion, SADN can still detect and locate it within a small range. When performing target detection on the RD map to be tested, the target trajectory is first obtained through MF-IFA-UNet matching. If any lost targets are then segmented, the resulting small-scale RD map of the local area is input into the SADN. Through this process, even if the target is not detected in a certain frame due to clutter interference, there is still a chance that it will be found again in the local range, effectively avoiding trajectory interruption and improving the stability and accuracy of multi-target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the method of Example 1 of the present invention.

[0038] Figure 2It is a flow chart of the overall detection network MF-IFA-Unet of the present invention.

[0039] Figure 3 It is a flow chart of the small area detection network SADN of the present invention.

[0040] Figure 4 2 is a schematic diagram of the detection rate results of the comparative experiment of Example 1 of the present invention.

[0041] Figure 5 2 is a schematic diagram of the average false alarm rate per frame results of the comparative experiment of Example 1 of the present invention.

[0042] Figure 6 2 is a schematic diagram of the average Euclidean distance results of the comparative experiment in Example 1 of the present invention.

[0043] Figure 7 2 is a schematic diagram of the multi-target tracking accuracy results of the comparative experiment of Example 1 of the present invention. DETAILED DESCRIPTION

[0044] A preferred embodiment of the present invention is further described below with reference to the accompanying drawings.

[0045] Example 1

[0046] See also Figure 1 As shown, this embodiment provides a method for detecting and tracking small targets on the sea surface based on multi-frame detection, comprising the following steps:

[0047] S1. Add a simulated target to the sea clutter background, then cut the data matrix after adding the simulated target into an M×P size (M is the number of range units, P is the number of Doppler frequency units). Then perform a fast Fourier transform and shift the frequency center to convert it into an RD map (range-doppler), i.e., a range-Doppler map, where range is the distance between the target and the radar, and doppler is the radial velocity of the target relative to the radar. The area in the RD map containing the simulated target is cropped to an N×N×3 size and set as a small area RD map.

[0048] In this embodiment, M = 342, P = 456, and N = 64. In more application scenarios, the size of the small area RD map is determined according to the maximum range of possible locations of the target.

[0049] S2. Preprocess the RD map and the small area RD map respectively. Specifically, perform channel compression on the color RD map and the small area RD map respectively and convert them into one-dimensional grayscale images to reduce the amount of calculation. Normalize the pixels in the RD map and the small area RD map, and normalize the pixels in the range of [0, 255] to the range of [0, 1].

[0050] S3. Design the overall detection network MF-IFA-UNet, input the preprocessed RD map into MF-IFA-UNet for training and save the weights; design the small area detection network SADN, input the preprocessed small area RD map into SADN for training and save the weights.

[0051] The present invention adopts a two-stage design for the small target detection process on the sea surface. In the first stage, an overall detection network MF-IFA-UNet (Multi-Frame Inter-Frame Attention UNet) is designed. The network is based on the UNet encoder-decoder architecture and is improved for small target detection on the sea surface.

[0052] Please refer to Figure 2 As shown in the figure, MF-IFA-UNet first stacks the RD graph data of k consecutive frames at the input layer to form a three-dimensional tensor of size M×P×k. The value of k is set based on the target motion characteristics, so that the pixel offset caused by the target motion within k frames does not exceed the target scale. The encoder uses three-dimensional convolution (3DConv) to extract spatiotemporal features: through multiple 3D convolutional layers and downsampling layers, the spatial dimension is gradually compressed and the number of channels is increased, thus refining a deep feature map that integrates information from multiple frames.

[0053] Taking into account that the temporal correlation of sea clutter is usually shorter than that of maneuvering targets (clutter speckles change faster in adjacent frames), the present invention embeds an inter-frame self-attention IFA (Inter-FrameAttention) mechanism in the encoder of MF-IFA-UNet. The IFA module processes the feature tensor output by a certain layer of the encoder along the time frame dimension: a compressed representation of the features of each frame is obtained through global pooling, and the attention weight distribution of each frame is obtained through convolution calculation, and then the original feature tensor is applied back to enhance the inter-frame correlation. The IFA module gives MF-IFA-UNet the ability to pay attention to target features at different times. For example, when the target is clear in the current frame but is partially submerged by clutter in the adjacent frame, the network can still capture the target based on the inter-frame correlation, thereby effectively reducing the missed detection rate and false detection rate of multi-frame detection.

[0054] In the UNet decoder, MF-IFA-UNet uses multiple levels of transposed convolution to gradually upsample feature maps to the same size as the input. It also uses skip connections to fuse features at the corresponding encoder scale with decoded features, compensating for the loss of detail during the encoding process. The resulting detection probability heatmap has the same size as a single-frame RD map, with the value of each pixel representing the probability of a moving target at that location. By extracting spatiotemporal features from multiple frames and fusing inter-frame attention, MF-IFA-UNet is able to highlight the energy trajectory of moving targets against strong sea clutter.

[0055] In this embodiment, k = 3, meaning three RD maps are grouped together and fed into the MF-IFA-UNet model for training and detection. When the target's Doppler frequency is outside the primary sea clutter band, the network can accurately detect small targets with an SCR as low as -16dB. Compared to traditional single-frame detection networks, the multi-frame detection model MF-IFA-UNet proposed in this paper achieves a higher detection probability at the same false alarm rate.

[0056] Please refer to Figure 3 As shown in the figure, to address the potential for missed target detection during target tracking, the present invention designs a Small Area Detection Network (SADN) to perform a secondary search for possible target locations. SADN accepts an N×N RD map as input, with N being 64 in this embodiment. SADN employs an encoder-decoder architecture, but compared to MF-IFA-UNet, SADN specifically enhances feature extraction capabilities.

[0057] The SADN encoder consists of three cascaded residual downsampling modules. Each Res Down module (Res Down) consists of two convolutional layers and a residual connection. After downsampling, the number of channels increases step by step (e.g., 64 → 128 → 256 → 512), and the spatial size decreases accordingly: N / 2 → N / 4 → N / 8. This Res Down module alleviates the vanishing gradient problem through constant shortcut connections, enabling efficient training of deep networks while preserving key target information.

[0058] Furthermore, the coordinate attention (CA) and efficient channel attention (ECABlock) mechanisms are introduced after each level of residual downsampling modules. The CA module fuses the spatial location and channel information of the feature map to enhance the network's perception of object shape and position. The ECA module assigns appropriate channel weights to the extracted features, highlighting feature channels related to the object and suppressing irrelevant channels without requiring a large number of additional parameters. After multiple layers of residual + attention extraction, the SADN encoder can obtain highly abstract deep features in the N / 8×N / 8×512 dimension.

[0059] The decoder part of SADN also adopts upsampling and skip connection structure. The CT transposed convolution module amplifies the feature map to generate deep features, and fuses them with the shallow features of the encoder through skip connections. The deep features are then sent to the DConv multi-layer convolution module to further extract feature information. Finally, the features are gradually restored to the original N×N size, and a target probability map of the same size as the input area is output.

[0060] During training, SADN and MF-IFA-UNet employ a similar cross-entropy loss, comparing the detection output with the ground-truth object within a small region at the pixel level to optimize network parameters. This design enables SADN to accurately detect faint targets within a local sub-image. Even when a target disappears from global detection due to clutter, SADN can still detect and locate it within a small region.

[0061] S4, receive radar signals and convert them into RD maps to be detected, input the accumulated k frames of RD maps to be detected into MF-IFA-UNet, extract the spatiotemporal features between frames, and output the detection probability heat map, whose pixel value represents the confidence of the target at the corresponding position; T The value is set to 1, which is less than μ T The value of is set to 0. After performing the clustering operation, the target center position is obtained. If the current tracking sequence of the target is empty, a new tracking sequence is initialized for each detection result;

[0062] S5. According to the target center position output by MF-IFA-UNet, the target trajectory prediction and matching mechanism is used for tracking between each frame. Specifically, the Kalman filter is used to predict the target's position in the current frame based on its historical state (position, speed). The predicted position is matched with the target position detected in the current frame, and the Mahalanobis distance Dist between each pair of predicted points and detected points is calculated as the cost matrix cost. The Hungarian algorithm is applied to solve the optimal match; the Mahalanobis distance Dist between the current frame and the predicted position is set within the preset threshold Dist. max The target within the predicted position is assigned to the corresponding track, and the target matching is judged to be successful; when no detected target falls into the neighborhood of the predicted position, the target is judged to be lost.

[0063] For successfully matched targets, it is considered that the target is tracked normally in the current frame and its trajectory status is updated; for unmatched predicted targets (possibly because the target is missed in the current frame detection), an N×N RD map centered on the predicted position is extracted.

[0064] S6. Input the N×N RD map obtained in S5 into the SADN, output the detection probability heat map, and perform clustering in the same manner as step S5 to determine whether there is a target in the local area that failed to be matched in S5. If the target is detected, it is re-included in the tracking sequence and the target trajectory is updated; otherwise, the matching fails, the trajectory is broken, and the number of consecutive lost frames is increased by 1. When three consecutive frames are lost, the trajectory is considered to have disappeared.

[0065] S7. Repeat steps S4 to S6 to process the RD images of subsequent consecutive frames in real time to achieve continuous and stable tracking of the small target on the sea surface until the radar signal ends.

[0066] In this example, the data containing the small moving target was cropped and spliced. Specifically, two consecutive 64-range-unit and 6,940-pulse units were extracted and spliced along the pulse dimension to create experimental sea clutter data with 64 range units and 13,880 pulse units. The target echo was simulated in the time domain based on the target and sea clutter model formulas and the target motion model formula. The target echo was added to the cropped sea clutter data at a specified SCR to obtain the total echo. Subsequently, 108 consecutive RD maps were obtained using a short-time Fourier transform (STFT) using 128 pulses.

[0067] In order to verify the robustness of the above method, in this embodiment, different sizes and speeds are set for small moving targets. In each set of echoes, the total radial length of the moving target is set to 5m. At the same time, various motion models with different initial velocity parameters, such as uniform speed, uniform acceleration and deceleration, are simulated; the target acceleration is -2m 2 / s to 2m 2 / s, and the initial velocity is between -10m / s and 10m / s.

[0068] During MF-IFA-UNet training, 10 sea clutter scenarios were used. This involved cropping 64 consecutive ranges from different locations as the sea clutter background. After adding simulated targets, 310 data sets were generated, each containing 108 consecutive frames of RD maps, for a total of 33,480 RD maps. The SCR was set between -20 and 10 dB, and each data set contained two targets. A 64×64 target region was cropped from this data to form the SADN training set. Sea clutter data without simulated targets was also added to the training set to improve model stability.

[0069] In the test data set, additional sea clutter data was used as the background, with simulated targets ranging from -20dB to 10dB, and a sampling interval of 1dB. In this example, the SCR, motion trajectory, and sea clutter background of the simulated targets in the test and training sets were identical. The MF-IFA-UNet and SADN models were then trained separately.

[0070] In order to verify the outstanding effect of the present invention, the DCNN (k=9) algorithm and a multi-strategy fusion tracking algorithm with an infrared small target segmentation network (LW_IRSTNet) as the detection head are used as controls in this embodiment. The obtained index test results are as follows: Figures 4 to 7 shown.

[0071] Please refer to Figure 4 As shown in Figure 2, the detection probability of both DCNN and this method is close to or equal to 100% when the SCR is high, but the detection probability of this method converges to 100% faster. Figure 5As shown in , since both this method and DCNN detect moving targets based on the idea of multiple frames, the average false alarm rate per frame is at a low level. Figure 6 and Figure 7 It can be seen that the IRSTNet-based model is based on a single-frame tracking operation. When crossing the sea peak, the target reconstruction ability is weak, which affects the tracking performance of the model. It is weaker than other methods in terms of average Euclidean distance (in pixels) and multi-target tracking accuracy (MOTA). At the same time, from the perspective of multi-target tracking accuracy, the tracking ability of DCNN is also weaker than this method.

[0072] The above comparative analysis shows that the proposed method outperforms the two existing methods mentioned above in many indicators, showing higher detection accuracy, lower false alarm rate, more precise target positioning and higher tracking reliability.

[0073] Example 2

[0074] This embodiment provides a small sea target detection and tracking system based on multi-frame detection corresponding to the method described in Example 1, including:

[0075] RD map generation module: used to add simulated targets in the sea clutter background, then cut the data matrix after adding the simulated targets into a set size and convert it into a simulated RD map, and cut the area containing the simulated targets in the simulated RD map into a set size to obtain a local RD map;

[0076] Preprocessing module: used to preprocess RD graph and local RD graph respectively;

[0077] Network training module: used to design the overall detection network MF-IFA-UNet, input the preprocessed simulated RD map into MF-IFA-UNet for training and save the weights; design the small area detection network SADN, input the preprocessed local RD map into SADN for training and save the weights;

[0078] Target localization module: This module is used to input the RD maps to be detected accumulated over several frames into MF-IFA-UNet, extract the spatiotemporal features between frames, and output a detection probability heat map. It then performs a clustering operation on the values judged to be true targets in the detection probability heat map to obtain the true target center position for each frame and output the detection result for each true target. If a true target appears for the first time, the current tracking sequence is empty, and a new tracking sequence is initialized for each detection result.

[0079] Target matching module: It is used to predict the position of the real target in the current frame based on the center position of the real target output by MF-IFA-UNet and the position and velocity of the real target in the previous frame using the Kalman filter. The detection result of the current frame is then matched with the predicted position. If the match is successful, the target trajectory is updated. If the match fails, the local RD map to be detected is cropped with the predicted position as the center.

[0080] Lost Target Matching Module: This module is used to input the local RD map to be detected obtained by the target matching module into the SADN, output a detection probability heat map, and perform a clustering operation to determine whether there is a real target in the local RD map to be detected that failed to be matched in the target matching module. If the real target is detected, it is re-included in the tracking sequence and the target trajectory is updated;

[0081] Continuous detection and tracking module: used to repeatedly execute the target positioning module, target matching module and lost target matching module in sequence, and detect and track the target frame by frame.

Claims

1. A method for detecting and tracking small targets on the sea surface based on multi-frame detection, characterized in that: The following steps are involved: S1. Add a simulated target to the sea clutter background, then cut the data matrix after adding the simulated target into a set size and convert it to obtain a simulated RD map, and cut the area containing the simulated target in the simulated RD map into a set size to obtain a local RD map; S2, preprocessing the simulated RD map and the local RD map respectively; S3. Design the overall detection network MF-IFA-UNet, input the preprocessed simulated RD graph into MF-IFA-UNet for training and save the weights; Design a small area detection network SADN, input the preprocessed local RD map into SADN for training and save the weights; S4. Input the RD maps to be detected accumulated from several frames into MF-IFA-UNet, extract the spatiotemporal features between frames, and output a detection probability heat map. Cluster the values judged as true targets in the detection probability heat map to obtain the true target center position in each frame, and output the detection result of each true target. If the true target appears for the first time, the current tracking sequence is empty, and a new tracking sequence is initialized for each detection result. S5. Based on the center position of the real target output by MF-IFA-UNet, the Kalman filter is used to predict the position of the real target in the current frame based on the position and velocity of the real target in the previous frame. The current frame detection result is then correlated and matched with the predicted position. If the match is successful, the target trajectory is updated. If the match fails, the predicted position is used as the center to crop the local RD map to be detected. S6. Input the local RD map to be detected obtained in S5 into the SADN, output the detection probability heat map, perform a clustering operation to determine whether there is a real target that failed to be matched in S5 in the local RD map to be detected. If the real target is detected, the real target is re-included in the tracking sequence and the target trajectory is updated; S7. Repeat steps S4 to S6 to detect and track the real target frame by frame.

2. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: In step S1, the model formula of sea clutter and simulated target is: Where: H0 is the model containing only sea clutter, H1 is the model containing simulated target echo and sea clutter, z(m,p) is the complex time series received by the radar, z p (m,p) is the sea clutter time series, τ(m,p) is the slowly varying non-negative texture sequence, u(m,p) is the rapidly varying speckle sequence, A0(m,p) is the complex coefficient amplitude matrix, j is the imaginary unit, R(p) is the radial distance of the target at the p-th pulse, λ is the radar wavelength, m is the m-th range unit, and p is the p-th pulse signal; The dynamic formula for simulating target motion is: Where: r (p) is the instantaneous radial velocity, a r (p) is the instantaneous acceleration, a r (p) obeys the normal distribution with mean μ r , with a standard deviation of δ r ; Δt is the sampling time.

3. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: In step S3, the overall detection network MF-IFA-UNet includes an encoder and a decoder. The encoder includes several convolutional layers and downsampling layers, and uses three-dimensional convolution to extract the joint features of the spatial and temporal dimensions of the input multi-frame RD graph; the decoder includes a multi-level transposed convolutional layer symmetrical to the encoder, and fuses the features of the corresponding convolutional layers of the encoder with the decoding features of the transposed convolutional layer through jump connections, and outputs a detection probability heat map with the same size as the input RD graph. The overall detection network MF-IFA-UNet embeds the inter-frame attention mechanism module IFA during the convolution process.

4. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: In step S3, the small area detection network SADN includes a downsampling encoder and an upsampling decoder. The downsampling encoder includes several levels of residual downsampling modules, each level of residual downsampling module includes two convolutional layers and a residual connection; the upsampling decoder includes a multi-level transposed convolutional layer symmetrical to the downsampling encoder, and fuses the features of the corresponding convolutional layer of the downsampling encoder with the decoded features of the transposed convolutional layer through jump connections, and outputs a detection probability heat map with the same size as the input simulated RD map.

5. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 4, characterized in that: In the downsampling encoder, the coordinate attention CA and efficient channel attention ECA mechanisms are introduced after each level of downsampling residual downsampling module.

6. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: In step S5, the Hungarian algorithm is used to associate and match the current frame detection result with the predicted position: The Mahalanobis distance between the center position of the true target in the detection result and the corresponding predicted position of the true target is calculated as the value of the cost matrix and assigned using the Hungarian algorithm; the true target whose Mahalanobis distance to the predicted position in the current frame is within the preset threshold is assigned to the corresponding track, and the target matching is determined to be successful; Otherwise, the real target is determined to be lost.

7. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: The step S6 comprises: A clustering operation is performed on the detection probability heat map to obtain the center position of the true target. When the true target is not empty, the Mahalanobis distance between the true target center position and the predicted position is calculated. When the Mahalanobis distance is less than the preset threshold, the target is judged to be matched successfully. Otherwise, the match fails and the frame is recorded as a lost frame. When the number of consecutive lost frames exceeds the set threshold, it is judged that the true target trajectory has disappeared.

8. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: In step S1, the generation process of the simulated RD graph is as follows: performing a sliding window operation on the data matrix after adding the simulated target, and then performing a fast Fourier transform and shifting the frequency center to generate the simulated RD graph.

9. The method for detecting and tracking small targets on the sea surface based on multi-frame detection according to claim 1, characterized in that: In step S2, the preprocessing includes: performing channel compression on the simulated RD map and the local RD map, converting them into one-dimensional grayscale images, and normalizing the pixels in the simulated RD map and the local RD map.

10. A small target detection and tracking system on the sea surface based on multi-frame detection, characterized in that: include: RD map generation module: used to add simulated targets in the sea clutter background, then cut the data matrix after adding the simulated targets into a set size and convert it into a simulated RD map, and cut the area containing the simulated targets in the simulated RD map into a set size to obtain a local RD map; Preprocessing module: used to preprocess RD graph and local RD graph respectively; Network training module: used to design the overall detection network MF-IFA-UNet, input the preprocessed simulated RD graph into MF-IFA-UNet for training and save the weights; Design a small area detection network SADN, input the preprocessed local RD map into SADN for training and save the weights; Target localization module: This module is used to input the RD maps to be detected accumulated over several frames into MF-IFA-UNet, extract the spatiotemporal features between frames, and output a detection probability heat map. It then performs a clustering operation on the values judged to be true targets in the detection probability heat map to obtain the true target center position for each frame and output the detection result for each true target. If a true target appears for the first time, the current tracking sequence is empty, and a new tracking sequence is initialized for each detection result. Target matching module: It is used to predict the position of the real target in the current frame based on the center position of the real target output by MF-IFA-UNet and the position and velocity of the real target in the previous frame using the Kalman filter. The detection result of the current frame is then matched with the predicted position. If the match is successful, the target trajectory is updated. If the match fails, the local RD map to be detected is cropped with the predicted position as the center. Lost Target Matching Module: This module is used to input the local RD map to be detected obtained by the target matching module into the SADN, output a detection probability heat map, and perform a clustering operation to determine whether there is a real target in the local RD map to be detected that failed to be matched in the target matching module. If the real target is detected, it is re-included in the tracking sequence and the target trajectory is updated; Continuous detection and tracking module: used to repeatedly execute the target positioning module, target matching module and lost target matching module in sequence, and detect and track the target frame by frame.

Images