A target detection method based on multi-path component reconstructed residual

Abstract

The application discloses a target detection method based on multi-path component reconstruction residual error, comprising the following steps: step 1: based on the double auto-encoder training algorithm of environmental clutter, learning the clutter structure features in the multi-path channel and reconstructing the multi-path echo; step 2: synthesizing the reconstructed multi-path echo obtained in step 1 into a reconstructed image, based on the target pre-detection algorithm of the reconstruction residual error, comparing the difference between the original image and the reconstructed image, and used for distinguishing whether the to-be-detected region contains target echo. The application fully explores the multi-path characteristics of the clutter, and realizes the weak and small target detection under the condition of no target sample through the multi-path echo reconstruction residual error.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to a target detection method based on multipath component reconstruction residuals. Background Technology

[0002] Radar detection suffers from the problem of difficulty in detecting small, non-cooperative targets. This is mainly because existing target detection methods require the target echo to be stronger than clutter, or rely on targeted echo samples. When the target is small or non-cooperative, the desired effect is often not achieved.

[0003] Effective detection of weak, non-cooperative targets in complex multipath aliasing environments involves three aspects: filtering out false targets from multipath paths, detecting weak targets, and suppressing noisy false alarm targets. For example... Figure 1 As shown, in complex multipath cascading situations in cities, the above-mentioned problems are often encountered when performing high-confidence detection of small targets such as drones.

[0004] Existing methods mainly fail to overcome the challenge of detecting weak targets in complex multipath aliasing environments. Among these challenges, the detection of weak, non-cooperative targets is particularly difficult. Targets such as drones and human bodies are small in size, have weak echoes, and move slowly, resulting in a low probability of target detection. Furthermore, the difficulty in obtaining samples leads to unknown echo characteristics.

[0005] Existing target detection methods based on target echo intensity can effectively filter out clutter when the target echo is stronger than the clutter, resulting in data containing only the target echo and good detection performance. However, when the target echo is weaker than the clutter, it becomes difficult to easily filter out the target echo from the clutter, leading to poor detection performance for small targets.

[0006] Existing machine learning-based object detection methods, which directly learn noise features and target echo features, can accurately identify targets similar to the learned samples. However, when samples are missing or the training samples do not match the measured targets, the detection performance degrades significantly. Furthermore, these methods cannot be used for non-cooperative target detection due to the lack of non-cooperative target echo samples. Summary of the Invention

[0007] In order to overcome the shortcomings of the existing technology, the present invention aims to provide a target detection method based on multipath component reconstruction residual. This detection method fully explores the multipath characteristics of clutter and reconstructs residual through multipath echo, so as to achieve weak target detection in the absence of target samples.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A target detection method based on multipath component reconstruction residuals includes the following steps;

[0010] Step 1: The dual autoencoder training algorithm based on environmental clutter learns the clutter structure characteristics in the multipath channel and reconstructs the multipath echo;

[0011] Step 2: Combine the reconstructed multipath echoes obtained in Step 1 into a reconstructed image. Based on the target pre-detection algorithm of the reconstruction residual, compare the differences between the original image and the reconstructed image to determine whether the area to be detected contains target echoes.

[0012] In step 1, the echo of environmental clutter is used as the training set. According to different propagation paths, the target echo obtained by electromagnetic waves is divided into direct wave component and multipath component, which are used as inputs to the dual autoencoder (the dual autoencoder requires multi-channel input, and the division is a necessary step. Although they are input together, the input channels are different, and using multipath component as multi-channel input can reflect the characteristics of this method in combining radar echo and machine learning); the dual autoencoder structure is encoder one - decoder - encoder two.

[0013] The optimal parameter matrices E, D, and F for encoder 1, decoder 2, and encoder 2 are found using machine learning methods to reconstruct the image q. ij and the original image p ij The closest reconstructed feature g ij and original feature f ij The closest result indicates that the dual autoencoder has completed the learning of clutter multipath echo characteristics.

[0014] That is, satisfy That is, by obtaining the parameters in encoders E and F and decoder D, the training image p can be optimized. ij With reconstructed image q ij The difference is less than 10%, and the echo characteristics are f ij With reconstruction feature g ij The similarity indicates that the dual autoencoder has successfully learned the characteristics of clutter multipath echoes.

[0015] This formula is derived from the fact that the reconstructed image and the original image are identical. It is used to train the parameters, finding the most suitable parameters that minimize the value of this formula. The optimal result is p. ij =q ij f ij =g ij That is, input equals output; p ij It is the input, f ij It is p ij After the result of the first encoder, q ij It is f ij g obtained after decoding ij It is q ij The values ​​α and β, obtained from the second encoder, are used to control the weights.

[0016] The dual autoencoder has multiple input channels, each of which inputs a multipath component. The multiple channels refer to the input channels of the dual autoencoder. Here, the multipath component features from multiple channels are fused to learn the relationship between multipath images.

[0017] Fusion refers to combining the structures obtained from each channel into a single coordinate system to form a reconstructed image. The dual autoencoder's function is that during training, the output equals the input, indicating training is complete; during use, the input is the same as the training input.

[0018] The encoders consist of a convolutional layer and a pooling layer, which compress and reduce the dimensionality of the input data to obtain its features. The decoder has the opposite structure to the encoder, consisting of a pooling layer and a convolutional layer, which decompresses the data features and reconstructs the output data similar to the input data. The reconstructed image features are then obtained through the encoder.

[0019] Step 1 specifically involves:

[0020] Step (1): When the radar acquires data, a multipath aliasing echo is obtained through the multipath effect, which is the original image. Multipath dealiasing is then used to decompose the original image into images p corresponding to the multipath components. ij .

[0021] Step (2): Convert the multipath component corresponding image p ij The input is given to encoder one, and then processed by encoder one E. ij Convolutional and pooling layers are used to obtain signal features f. ij ;

[0022] Step (3): Transfer the signal feature f ij Input to decoder D ij Reconstruct the image q ij ;

[0023] Step (4): Then process the reconstructed image q ij via encoder 2F ij Encoding yields the reconstructed feature g i j .

[0024] In step 2, a target pre-detection algorithm is used to reconstruct residuals using multipath components to distinguish between targets and clutter, thereby achieving detection of the area to be detected.

[0025] Once the encoder training is complete, matrices E, D, and F have learned the clutter multipath characteristics. At this point, the reconstruction residuals of the target area can be used to... To determine whether a target echo exists.

[0026] After reconstructing the multipath components input in each channel using matrices E, D, and F, the reconstructed multipath components are similar to the input image, and the residuals of each channel are small, indicating that the features of the region to be detected conform to clutter features, i.e., the region to be detected is not a target. Conversely, when the region to be detected contains a target, the multipath components conform to the target features, which are different from the clutter features in matrices E, D, and F. Therefore, the reconstructed multipath components of each channel differ significantly from the input target multipath components, and the residual of the reconstructed image will increase significantly. Thus, the signal feature detection of non-cooperative targets can be achieved by reconstructing the residuals.

[0027] The input image is divided into two types:

[0028] The first type is an image of a cluttered environment without a target, which is used to train a dual autoencoder so that the reconstructed image is basically the same as the original image, thus completing parameter training.

[0029] The second type is the image to be inspected, which is where it is unknown whether it contains a target. Using the dual autoencoder trained in the first step, a reconstructed image is obtained. If the image is basically the same as the original image, it means there is no target; if it is significantly different, it means there is a target.

[0030] The beneficial effects of this invention are:

[0031] The key to this invention is the proposal of a target detection method based on multipath component reconstruction residuals. The method trains dual encoders to fully exploit the characteristics of environmental clutter through multipath echoes for unsupervised learning. By relying on the large differences in the reconstruction of target multipath components and the small differences in the reconstruction of clutter multipath components, the method achieves the detection of weak targets without samples.

[0032] First, this invention has better target detection capabilities. Existing methods rely on the target echo intensity, making them unsuitable for environments where the target echo intensity is lower than the clutter intensity. This invention utilizes the echo characteristics of both the target and clutter multipath components, achieving better results even in environments where the target echo intensity is lower than the clutter intensity, and enabling the detection of small targets.

[0033] Secondly, this invention has strong practicality. Existing methods require learning samples of the target, and the lack of learning samples or incorrect samples will greatly affect the detection performance. This invention uses environmental clutter multipath components as samples for learning, eliminating the need to detect target samples. Moreover, environmental samples are easier to obtain than target samples, enabling the detection of non-cooperative targets without available samples. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating high-confidence detection of small, unmanned aerial vehicle (UAV) targets in complex, multipath-overlapping urban environments.

[0035] Figure 2 This is the set of assumptions about the multipath propagation paths of electromagnetic waves.

[0036] Figure 3 This is a schematic diagram of a target detector based on multipath components and dual autoencoders.

[0037] Figure 4 This diagram illustrates the situation where the area to be inspected is clutter, the reconstruction residual is small, and the detection is clutter.

[0038] Figure 5 This diagram illustrates the presence of a target echo in the area to be inspected, the large reconstruction residual, and the detection of the target. Detailed Implementation

[0039] The present invention will now be described in further detail with reference to the accompanying drawings.

[0040] This invention provides a target detection method based on multipath component reconstruction residuals, which is generally divided into two parts: a dual autoencoder training algorithm based on environmental clutter and a target pre-detection algorithm based on reconstruction residuals.

[0041] In principle, by performing unsupervised learning on clutter multipath features, an encoder is trained to record clutter multipath features. Taking advantage of the encoder's characteristic of low reconstruction residuals for clutter multipath components and high reconstruction residuals for target echo multipath components, clutter and targets are distinguished by the reconstruction residuals of multipath components. The probability of detecting weak and non-cooperative targets is improved by utilizing echo features, so as to achieve effective detection of weak targets when target samples are scarce.

[0042] In complex aliased multipath environments, due to multipath effects, the radar's target detection echo consists of direct echoes and multipath echoes. If only the cases of electromagnetic wave reflections at two times or less are considered, filtering out weaker higher-order multipath reflections is necessary. Figure 2 As shown, electromagnetic waves propagate twice in the environment, once for transmission and once for reception. During transmission, there are three possible reflections: 0, 1, and 2. Similarly, during reception, there are a total of 3×3 paths. Therefore, there are 9 possible paths for the echo.

[0043] Multipath echoes are not entirely meaningless interference data; each multipath echo contains a portion of the target's features. Multipath echoes are identical images produced by the multipath effect of target and environmental echoes, only arriving at different times. Theoretically, they contain all features, but due to factors such as echo intensity and environmental interference, some features are randomly lost. However, the loss from each multipath is different. By fusing multiple multipaths, the mutual losses can be compensated for, making the target's features more apparent.

[0044] If the obtained multipath echoes are utilized, the understanding of target characteristics will be more comprehensive. Therefore, this invention uses unsupervised learning to learn the multipath features of easily obtainable clutter. That is, the multipath echo data obtained from these 9 paths are used as input to train a dual autoencoder to learn the multipath features of clutter. Then, the residuals are reconstructed through multipath components to determine whether there is a target in the region, thereby completing the detection of weak targets.

[0045] This invention was developed for detecting weak, non-cooperative targets and is applicable to radar detection and imaging of weak targets such as drones and human bodies. Because this invention utilizes multipath components to reconstruct radar images, it is also suitable for complex aliased multipath environments with multiple targets, multiple reflective surfaces, and multiple reflections overlapping.

[0046] A target detection method based on multipath component reconstruction residuals includes the following steps:

[0047] Step 1: Dual autoencoder training algorithm based on environmental clutter:

[0048] This invention utilizes environmental clutter to train a dual autoencoder for unsupervised learning. Learning is performed using current clutter data, and the encoder and decoder are trained to learn the clutter structure features within the multipath channels based on multipath echoes under various multipath assumptions. After training, when clutter is input to the multipath channel (i.e., the input channel of the dual autoencoder; since the input in this paper refers to the various multipath components), the encoder and decoder can reconstruct the signal characteristics of the clutter well, with a small residual in the reconstructed multipath signal. However, when the input is a target echo, the difference between the multipath echo and the reconstructed multipath echo is significant; the reconstructed residual can be used to determine whether the area to be inspected contains the target echo.

[0049] Figure 3 This is a schematic diagram of a target detector based on multipath components and dual self-encoders. The direct wave component and multipath component are electromagnetic waves based on... Figure 2 The target echo obtained through different propagation paths is used as the input of the dual autoencoder; the dual autoencoder is obtained by adding an encoder on the basis of the ordinary autoencoder, and the structure is encoder-decoder-encoder.

[0050] Using the echoes of environmental clutter as the training set, machine learning methods are used to find the optimal parameter matrices E, D, and F of the encoder, such that the reconstructed image q... ij and the original image p ij The closest reconstructed feature g ij and original feature f ij Closest, i.e., satisfies

[0051] The specific training process for the encoder-decoder is as follows:

[0052] Step 1: Convert the multipath component corresponding image p ij The input is given to the encoder, and then transmitted through encoder E. ij Convolutional and pooling layers are used to obtain signal features f. ij ;

[0053] Step 2: Extract signal features f ij Input to decoder D ij It is possible to reconstruct the image q ij ;

[0054] Step 3: Then process the reconstructed image q ij via the second encoder F ij Encoding yields the reconstructed feature g ij .

[0055] Unlike traditional autoencoders, this method fuses multipath component features from multiple channels to learn the relationships between multipath images.

[0056] The training process can be represented as That is, obtain encoder E ij and F ij Decoder D ij The parameters in the training image p can make the training image p ij With reconstructed image q ij Very similar, echo characteristics f ij With reconstruction feature g ij The similarity indicates that the dual autoencoder has successfully learned the characteristics of clutter multipath echoes.

[0057] An autoencoder is an unsupervised learning algorithm that aims to generate output data similar to the input data by learning the inherent structure and features of the input data.

[0058] The dual self-encoder structure consists of encoder-decoder-encoder.

[0059] The encoders consist of a convolutional layer and a pooling layer, which compress and reduce the dimensionality of the input data to obtain its features. The decoder has the opposite structure to the encoder, consisting of a pooling layer and a convolutional layer, which decompresses the data features and reconstructs the output data similar to the input data. The output data is then passed through a second encoder layer to obtain the features of the reconstructed image.

[0060] During parameter training, the sum of squares of feature differences and reconstructed image differences is used as an indicator to complete the parameter training process, that is, to complete the dual autoencoder's full learning of the echo characteristics of the clutter environment.

[0061] Figure 3 In this context, matrices E, D, and F are merely parametric matrices; their values ​​are obtained from the training process and have no real meaning. They are only used to achieve [the desired result]. The purpose of this formula.

[0062] Step 2: Target pre-detection algorithm based on reconstruction residuals

[0063] The target pre-detection method of the present invention utilizes the reconstruction residual of multipath components to distinguish between the target and the echo, thereby realizing the detection of the area to be detected.

[0064] Once the encoder training is complete, matrices E, D, and F have learned the clutter multipath characteristics. At this point, the reconstructed residuals of the target area can be used to... To determine whether a target echo exists.

[0065] like Figure 4 As shown, the multipath components input in each channel are reconstructed using matrices E, D, and F. The reconstructed multipath components are similar to the input image, and the residuals of each channel are small, indicating that the features of the region to be detected conform to clutter features, that is, the region to be detected is not a target.

[0066] Conversely, if Figure 5 As shown, when the region to be detected contains a target, the multipath components conform to the target characteristics, which are distinct from the clutter characteristics in matrices E, D, and F. Therefore, the reconstructed multipath components differ significantly from the input target multipath components, resulting in a substantial increase in the reconstructed image residual. Thus, signal feature detection of non-cooperative targets can be achieved through the reconstruction residual. Because parameter matrices E, D, and F are trained using a clutter environment, the reconstructed image is identical to the original image only when the input is environmental clutter. Therefore, when a target is present, the input differs from the training input, leading to significant differences in the reconstructed image. Figure 4 and Figure 5 This is a display of results, showing small and large differences.

[0067] The key to this invention is the proposal of a target detection method based on multipath component reconstruction residuals. The method trains dual encoders to fully exploit the characteristics of environmental clutter through multipath echoes for unsupervised learning. By relying on the large differences in the reconstruction of target multipath components and the small differences in the reconstruction of clutter multipath components, the method achieves the detection of weak targets without samples.

[0068] The absence of samples is reflected in the fact that the training of the dual autoencoder did not use data containing target samples, because target samples are sometimes difficult to obtain; instead, it used target-free environmental samples (i.e., clutter samples), which are easy to obtain. "Weak" refers to low target echo intensity. This method does not use echo intensity to determine the presence of a target, so it can detect weak targets.

Claims

1. A target detection method based on multipath component reconstruction residuals, characterized in that, Includes the following steps; Step 1: The dual autoencoder training algorithm based on environmental clutter learns the clutter structure characteristics in the multipath channel and reconstructs the multipath echo; Step 2: Combine the reconstructed multipath echoes obtained in Step 1 into a reconstructed image. Based on the target pre-detection algorithm of the reconstruction residual, compare the differences between the original image and the reconstructed image to determine whether the area to be detected contains target echoes. In step 1, the echo of environmental clutter is used as the training set. According to different propagation paths, the target echo obtained by electromagnetic waves is divided into direct wave component and multipath component, which are used as the input of the dual autoencoder. The dual autoencoder structure is encoder one - decoder - encoder two. The input channels of a dual self encoder include multiple channels, each channel inputting a multipath component. Multiple channels refer to the input channels of the dual self encoder. This section fuses the multipath component features from multiple channels to learn the relationships between multipath images; Fusion refers to putting the structures obtained from each channel together in the same coordinate system to form a reconstructed image; the function of the dual autoencoder is that the output equals the input during training, and the input is the same as the training input when in use; Step 1 specifically involves: Step (1): When the radar acquires data, a multipath aliasing echo is obtained through the multipath effect, which is the original image. Multipath dealiasing is used to decompose the original image into multipath components corresponding to the original image p. ij ; Step (2): Assign the multipath components to the original image p ij The input is given to the encoder, and then transmitted through encoder E. ij Convolutional and pooling layers are used to obtain signal features f. ij ; Step (3): Transfer the signal feature f ij Input to decoder D ij Reconstruct the reconstructed image q ij ; Step (4): Then process the reconstructed image q ij via the second encoder F ij Encoding yields the reconstructed feature g ij .

2. The target detection method based on multipath component reconstruction residuals according to claim 1, characterized in that, The optimal parameter matrices E, D, and F for encoder 1, decoder 2, and encoder 2 are found using machine learning methods to reconstruct the image q. ij and the original image p ij The closest reconstructed feature g ij and original feature f ij The closest, i.e., the encoder E is obtained. ij and F ij Decoder D ij The parameters in the image can make the original image p ij With reconstructed image q ij Very similar, echo characteristics f ij With reconstructed feature g ij The similarity indicates that the dual autoencoder has successfully learned the characteristics of clutter multipath echoes. That is, satisfy ; This formula is derived from the fact that the reconstructed image and the original image are identical. It is used for parameter training to find the most suitable parameters that minimize the value of this formula. The optimal result is p. ij =q ij f ij =g ij That is, input equals output; p ij It is the input, f ij It is p ij After the result of the first encoder, q ij It is f ij g obtained after decoding ij It is q ij The values ​​α and β, obtained from the second encoder, are used to control the weights.

3. The target detection method based on multipath component reconstruction residuals according to claim 1, characterized in that, Each encoder consists of a convolutional layer and a pooling layer, which compresses and reduces the dimensionality of the input data to obtain the features of the input data. The decoder is the opposite structure to the encoder. It consists of a pooling layer and a convolutional layer. It decompresses the data features and reconstructs the output data to be similar to the input data. The output data is then passed through the encoder to obtain the features of the reconstructed image.

4. The target detection method based on multipath component reconstruction residuals according to claim 1, characterized in that, In step 2, a target pre-detection algorithm for reconstructing residuals using multipath components is used to distinguish between targets and clutter, thereby achieving detection of the area to be detected. Once the encoder training is complete, matrices E, D, and F have learned the clutter multipath characteristics. At this point, the reconstruction residuals of the target area can be used to... This is used to determine whether a target echo exists.

5. The target detection method based on multipath component reconstruction residuals according to claim 4, characterized in that, After reconstructing the multipath components input in each channel using matrices E, D, and F, the reconstructed multipath components are similar to the input image, and the residuals of each channel are small, indicating that the features of the region to be detected conform to clutter features, i.e., the region to be detected is not a target. Conversely, when the region to be detected contains a target, the multipath components conform to the target features, and the target features are different from the clutter features in matrices E, D, and F. Therefore, the reconstructed multipath components of each channel differ greatly from the input target multipath components, and the residual of the reconstructed image will increase significantly. Thus, the signal feature detection of non-cooperative targets can be achieved by reconstructing the residuals.

6. The target detection method based on multipath component reconstruction residuals according to claim 5, characterized in that, The input image is divided into two types: The first type is an image of a cluttered environment without a target, which is used to train a dual autoencoder so that the reconstructed image is basically the same as the original image, thus completing parameter training. The second type is the image to be inspected, which is where it is unknown whether it contains a target. Using the dual autoencoder trained in the first step, a reconstructed image is obtained. If the image is basically the same as the original image, it means there is no target; if it is significantly different, it means there is a target.

Images