An ICRN-based sequential UWB-IR image vehicle target identification method

Abstract

The application discloses a kind of sequential UWB-IR image vehicle target identification methods based on ICRN, including steps S100: UWB-IR vehicle target imaging dataset construction based on array imaging method;Step S200: ICRN network model design, including three sub-modules of mask layer, iteratively convolution layer and RNN layer, the model utilizes single image depth feature and the depth correlation feature between sequence images, and has dynamic self-adaptability to sequential image number;Step S300: ICRN model classification performance verification based on UWB-IR simulation vehicle image set;Step S400: generalization and dynamic sequential target identification performance analysis of ICRN model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of UWB-IR target recognition, and particularly relates to a sequential UWB-IR image vehicle target recognition method based on ICRN. BACKGROUND

[0002] Ultra Wide-band Radar (UWBR) is a new type of radar technology, which has been widely used in gesture recognition, vital sign detection, human imaging or automatic driving and other military and civilian fields, and has achieved good application effect. According to the different waveforms of the transmitted signals, UWB can be divided into two categories: impulse UWB (UWB-IR) and non-impulse UWB. UWB-IR, also known as carrierless radar, non-sinusoidal wave radar, baseband radar, etc., is a typical transient radar, and its radiation usually has a sequence of impulse pulses with ultra-wide bandwidth. Most of the current commonly used UWB-IR target recognition methods are based on one-dimensional echo signals of the target, i.e. high resolution range profile (HRRP). Common HRRP-based target recognition methods include traditional HRRP feature extraction and analysis methods, HRRP two-dimensional processing methods, and deep learning-based HRRP target recognition methods. In addition to HRRP target recognition methods, array UWB-IR can also be used to perform two-dimensional imaging of the target, and target recognition can be performed based on the target image. It is worth noting that the UWB-IR image target is a distance-angle image of the target, and the recognition process is different from the classical SAR image recognition.

[0003] In fact, in addition to the above-mentioned recognition method based on the features of a single image, the performance of target detection, positioning and recognition under complex scene conditions can also be improved by the continuous change rule or correlation feature of the image in a certain dimension (such as time dimension). For example, in the application scenarios of behavior detection based on video stream, trajectory prediction, and emotion recognition based on continuous speech and text, the above-mentioned target recognition method based on correlation features has achieved relatively obvious effect. In the current commonly used UWB-IR target recognition methods, such as target HRRP recognition, the target distance information feature is used to realize target recognition. This method is greatly affected by the observation angle, and only contains the distance dimension information of the target. Some other methods use imaging algorithms to construct the imaging graph of the target relative to the radar, and realize the fusion of the distance and angle dimension information of the target. This method improves the information quantity of the data, and then obtains more effective features. However, even if the information is fused, the result of this method will also be greatly affected when the distance-angle information of the target changes greatly. In addition, in the actual test scene, with the complex environmental setting and possible interference, UWB-IR target recognition under low signal-to-noise ratio is also a difficult task.

[0004] On the basis of the above research, the patent proposes a UWB-IR sequential target recognition method under low signal-to-noise ratio, which uses array antenna imaging for UWB radar imaging, extracts the continuous features between images from the continuous image signals collected by the radar, and designs a neural network model ICRN (Iterable Convolution Recurrent Network, ICRN), proposes a UWB-IR sequential target recognition method based on ICRN, and finally realizes the recognition of sequential targets. SUMMARY

[0005] In order to overcome the deficiencies in the prior art, the present application provides a sequential UWB-IR image vehicle target recognition method based on ICRN, which can realize UWB-IR target recognition task by using continuous features of targets under low signal-to-noise ratio, and can provide more detailed information for UWB-IR target recognition in complex environment.

[0006] In order to achieve the above invention purpose and solve its technical problems, the technical scheme adopted is as follows:

[0007] A sequential UWB-IR image vehicle target recognition method based on ICRN, comprising the following steps:

[0008] Step S100: UWB-IR vehicle target imaging data set construction based on array imaging method: introduce array imaging algorithm into UWB-IR, give UWB-IR target imaging method, and simulate UWB-IR image of vehicle target under different observation conditions according to given vehicle scattering point model, and construct UWB-IR vehicle data set;

[0009] Step S200: ICRN network model design: design ICRN model, including mask layer, iteratively convolution layer and RNN layer three submodules, ICRN model uses single image target depth feature and sequence image target depth correlation feature, and has dynamic self-adaptability to sequence image number;

[0010] Step S300: ICRN model effect verification: verify the effect of ICRN model on the simulation generated UWB-IR vehicle data set, and compare with other deep network methods;

[0011] Step S400: model generalization and dynamic sequential target recognition performance: use different distribution data of the same target to train and verify ICRN, analyze the generalization of ICRN according to the performance of the verification set; at the same time, analyze the dynamic sequential UWB-IR vehicle target recognition performance of the model when the radar and the target exist relative motion.

[0012] Further, in step S100, the UWB-IR vehicle target imaging method introduces array imaging into UWB-IR, and gives a UWB-IR target imaging method; according to a given vehicle scattering point model, UWB-IR images of a vehicle target under different observation conditions, including a relative displacement distance of the target from a radar, a relative displacement angle, an observation azimuth angle and noise influence, are simulated, and a UWB-IR vehicle dataset is constructed.

[0013] Further, in step S200, the ICRN network model includes a Masking module, an iterative convolution module and an RNN module, and each module specifically includes:

[0014] The Masking module estimates dhash calculation of an input image and a reference image, sets a dhash threshold of sequential image selection, and filters and encodes an image sequence;

[0015] The iterative convolution module includes four layers of convolution, and the convolution kernel sizes are 5*5, 3*3, 3*3 and 3*3 respectively, the input and output channels are both 1, each layer of convolution is attached to a pooling layer and a batch_normalization layer with a size of 3*3 and a step of 2; in addition to the basic four-layer convolution, an update weight group for updating a subsequent picture is included, and the parameter dimension is consistent with the basic four-layer convolution model parameter dimension, that is, ω k+1 =ω k ·ω iterable +b iterable , wherein ω iterable and b iterable are update weights, ω k is a convolution parameter of a previous picture, and ω k+1 is a convolution parameter of a subsequent picture;

[0016] The RNN module includes an LSTM layer with 64 nodes, two fully connected layers, and the input and output are 64*64 and 64*3 respectively;

[0017] The output of the last fully connected layer does not contain an activation function, and the selection of all other activation functions contains a ReLU, and the model loss function selects a cross-entropy function.

[0018] Further, for the Masking module, k pieces of data continuous in the time dimension are input data, marked as t1 to t k After filtering and filtering of the input data by the Masking module, the first m sequential images after filtering are taken, or zero padding is performed to m pictures, which are transmitted to the iterative convolution module, and the number of valid sequential images is marked;

[0019] For the said iterable convolution module, it respectively carries out convolution feature extraction on the time dimension continuous data transmitted by the Masking module, the convolution parameters satisfy the transmission relationship, and the extracted features are spliced into a time sequence according to the time dimension and transmitted to the RNN module;

[0020] For the said RNN module, the LSTM model is used to further extract features of the input feature sequence, and according to the sequence effective length provided by the Masking module, the model update is stopped after the iterative output reaches the actual length of the sequence; finally, the output features are classified by using two full connection layers, and the classification results are output.

[0021] Further, the ICRN model uses the Masking module to ensure the continuity and effectiveness of the input data in the time dimension; the ICRN model uses the iterable convolution module to preliminarily extract the continuous features of the continuous data, and at the same time, when processing input of any length, the ICRN model does not increase the parameters; the ICRN model further extracts and utilizes the target continuous features by fusing the iterable convolution and the RNN module.

[0022] Further, in step S300, the ICRN model effect verification, in step S100 generated simulation data, respectively using ICRN model and other deep network method, comparative experiment effect and analysis.

[0023] Further, in step S400, the model generalization and dynamic sequential target recognition performance, the implementation process comprises:

[0024] For the model generalization, arrange the data according to the observation condition, select a part of the observation condition data for model training, use the remaining part to verify, and analyze the generalization recognition performance;

[0025] For the dynamic sequential target recognition performance of the model, the offset distance, offset angle, observation azimuth angle and signal-to-noise ratio of the data are controlled to be unchanged, and the dynamic sequential UWB-IR vehicle target recognition performance of the model is recorded as the detection distance approaches.

[0026] The above technical scheme is adopted in the present application, compared with the prior art, and has the following advantages and positive effects:

[0027] The present application proposes a sequential UWB-IR image vehicle target recognition method based on ICRN, which fully utilizes the single image target depth feature and the depth correlation feature between the sequence image targets, simultaneously considers the online adaptability of the number of sequence images, proposes an ICRN model, and realizes efficient recognition of UWB-IR sequential targets under low signal-to-noise ratio, providing a new technical approach for UWB-IR target recognition in complex environment. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the following embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor. In the drawings:

[0029] Figure 1 The flowchart of a sequential UWB-IR image vehicle target recognition method based on ICRN of the present application is shown in

[0030] Figure 2 Three vehicle target scattering models are shown in (a) off-road vehicle, (b) special vehicle, and (c) van truck.

[0031] Figure 3 The observation geometric position relationship diagram of the array UWB-IR radar to the vehicle target is shown in

[0032] Figure 4 The UWB-IR vehicle target imaging result diagram based on the array imaging method is shown in

[0033] Figure 5 The training and verification results of the ICRN model under the condition of PD=0m and SNR=inf are shown in (a) training set loss curve, (b) verification set loss function, (c) training set accuracy curve, and (d) verification set accuracy curve.

[0034] Figure 6 The dynamic sequential classification performance of the ICRN model under the condition of PD=0m and SNR=-10dB is shown in DETAILED DESCRIPTION

[0035] The technical solutions of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0036] The present embodiment discloses a sequential UWB-IR image vehicle target recognition method based on ICRN, and the flowchart is shown in Figure 1 The method comprises the following steps:

[0037] Step S100: Construction of UWB-IR vehicle target imaging data set based on array imaging method: introduce the array imaging algorithm into UWB-IR, give the UWB-IR target imaging method, and simulate the UWB-IR image of the vehicle target under different observation conditions according to the given vehicle scattering point model, and construct the UWB-IR vehicle data set.

[0038] Step S200: ICRN network model design: design the ICRN model, including three sub-modules of mask layer, iteratively convolutional layer and RNN layer, the ICRN model makes full use of the depth features of a single image target and the depth correlation features between sequence image targets, and has dynamic adaptability to the number of sequence images;

[0039] Step S300: ICRN model effect verification: verify the effect of the ICRN model on the UWB-IR vehicle data set generated by simulation, and compare it with other deep network methods;

[0040] Step S400: model generalization and dynamic sequential target recognition performance: use different distribution data of the same target to train and verify the ICRN, analyze the generalization of the ICRN according to the performance of the verification set; at the same time, analyze the dynamic sequential UWB-IR vehicle target recognition performance of the model when the radar and the target exist relative motion.

[0041] The above steps will be described in detail as follows:

[0042] In step S100, the UWB-IR vehicle target imaging data set is a distributed target model expanded based on a target strong scattering point model, and the imaging data of the target is collected by comprehensively considering the offset distance, offset angle, observation azimuth angle and noise influence of the target relative to the radar.

[0043] Preferably, in step S100, the vehicle target model constructed by the data set is as shown in Figure 2 The related parameter settings of the target are shown in Table 1. Figure 2 In the embodiment, the target model is composed of a plurality of dense points, wherein the intensity of each point is divided into three grades of strong, medium and weak, and the basic parameters of the target model are introduced in Table 1.

[0044] Table 1 shows the related parameter settings of the vehicle target model.

[0045]

[0046] Table 2 shows the related parameter settings of the array radar used for imaging.

[0047] Table 2 shows the parameter settings of the radar.

[0048] Parameter name Setting Signal sampling rate 100 GHz Signal bandwidth 1.6 GHz Transmit signal period 1 μs Number of transmit antennas / inter-element spacing 5 / 5 cm Number of receive antennas / inter-element spacing 5 / 5 cm

[0049] Further, in actual data collection, due to the complex observation conditions existing objectively, Figure 3 The possible situations of the array UWB-IR radar when observing the target are given. Figure 3The default target ground is XOY plane, the intersection of radar LOS and the plane is coordinate origin O, set the direction of radar as X positive direction, and establish the coordinate system as shown in the figure. In the above figure, R represents radar, T represents target position, set the distance from radar to O as D, set the distance from the center of T to O as PD (pre-dis, PD), set the angle between OT direction and OY direction as θ pa , called PA (pre-angle, PA), set the angle between the direction of target and OX direction as θ sa (switch-angle, SA). According to the above setting, Figure 3 The multiple observation conditions that may occur when the radar detects the target are simulated, and according to the above setting, the imaging results of the same target by the radar under different observation conditions will also be very different.

[0050] In summary, the variables D, PD, PA, SA and background noise SNR are controlled respectively, and on the basis of comprehensively considering various observation conditions, the UWB-IR target imaging data set is obtained. In the simulation data, the target offset distance is set to 0, 3, 5 m respectively, and when the offset distance is not 0, an offset angle is selected every 20°, at the same time, the target observation azimuth angle is selected every 10°, in each observation condition, the radar detection distance is set from 90 m to 10 m, and the sampling interval is 1 m, finally, in order to make the data more realistic, noise with a signal-to-noise ratio of-10 db is added. In summary, the data distribution is shown in Table 3.

[0051] Table 3 UWB-IR vehicle target imaging data set statistics

[0052]

[0053] Figure 4 The effect schematic diagram of several groups of imaging is shown, wherein the position of the radar is on the left side of the imaging diagram, the horizontal direction of the imaging diagram is distance, and the vertical direction is angle.

[0054] In step S200, an ICRN (Iterable Convolution Recurrent Network, ICRN) model is designed. Specifically, the overall structure of the model is as shown in Figure 1 The ICRN network model includes a Masking module, an iterable convolution module (IConvM) and an RNN module, and each module specifically includes:

[0055] The Masking module estimates the dhash calculation of the input image and the reference image, sets the dhash threshold of the sequential image selection, filters and encodes the image sequence;

[0056] The iterable convolution module (IConvM) includes four layers of convolution, and the convolution kernel sizes are 5*5, 3*3, 3*3 and 3*3 respectively, the input and output channels are both 1, each layer of convolution is attached with a pooling layer and a batch_normalization layer with a size of 3*3 and a step of 2; in addition to the basic four-layer convolution, it also includes a set of update weights for updating subsequent pictures, and the parameter dimensions are consistent with the basic four-layer convolution model parameter dimensions, that is, ω k+1 =ω k ·ω iterable +b iterable , wherein ω iterable and b iterable are update weights, ω k is the convolution parameter of the previous image, and ω k+1 is the convolution parameter of the next image.

[0057] The RNN module includes one LSTM layer with 64 nodes, two fully connected layers, and the input and output are 64*64 and 64*3 respectively.

[0058] The output of the last fully connected layer does not contain an activation function, and the selection of all other activation functions contains ReLU, and the model loss function is selected as the cross-entropy function; the setting of the above model parameters can be adjusted according to the experimental results.

[0059] Further, for the Masking module, the input data is k consecutive data (images) in the time dimension, marked as t1 to t k After filtering the input data by the Masking module, the first m sequential images are taken or zero-padded to m images, which are transmitted to the iterable convolution module IConvM, and the number of valid sequence images is marked.

[0060] For the iterable convolution module, the time-dimension continuous data transmitted by the Masking module are respectively subjected to convolution feature extraction, and the convolution parameters satisfy the transmission relationship, and the extracted features are spliced into a time sequence according to the time dimension and transmitted to the RNN module.

[0061] For the RNN module, the LSTM model is used to further extract features from the input feature sequence, and according to the sequence length provided by the Masking module, the model update is stopped after the actual length of the sequence is reached. Finally, the output features are classified using two fully connected layers, and the classification results are output.

[0062] Further, the ICRN model uses a masking module to ensure the continuity and effectiveness of the input data in the time dimension; the ICRN model uses an iterative convolution module to preliminarily extract the continuous features of the continuous data, and at the same time, the ICRN model does not increase the parameters when processing input of any length; the ICRN model further extracts and utilizes the target continuous features by fusing the iterative convolution and the RNN module.

[0063] In step S300, the ICRN model effect verification, in step S100 generated simulation data, respectively using ICRN model and other deep network method, comparative experiment effect and analysis.

[0064] In this embodiment, after the model design is completed, the UWB-IR vehicle target imaging data set in step S100 is used to verify the performance of the ICRN model. Specifically, 30 pictures with continuous observation distances under the same observation condition are combined into a sequence data, that is, k=30, and the maximum output length after filtering is set to 10, that is, m=10. In addition, considering that the data under the offset distance 0 condition is less than that under the other two offset distances, the dimension confusion is easy to form uneven sample distribution, so all experiments will be carried out under three offset distance conditions. The parameter settings of the ICRN model in this embodiment are shown in Table 4:

[0065] Table 4 ICRN model and training parameter settings

[0066]

[0067]

[0068] Figure 5 In order to control the PD to be 0m and not to add noise, the training and verification of the ICRN model, Table 5 lists the experimental results of different methods under various conditions.

[0069] Table 5 Experimental results of each method

[0070]

[0071] In the above table, it can be found that:

[0072] 1) With the increase of PD, the recognition performance of the three methods decreases, among which GoogleNet decreases relatively slowly, and CNN and ICRN decrease relatively obviously. Analysis shows that under the condition of large PD, with the change of PA, the target imaging result changes obviously, and the observation condition of the target by the radar changes dynamically, at this time, it is more difficult for CNN with relatively simple structure and ICRN focusing on the continuous change rule of the target.

[0073] 2) In the case of no noise, the performance of GoogleNet is better than CNN and ICRN. Analysis shows that: in the case of no noise, the difference between classes of UWB-IR vehicle target imaging data is large, and the feature extraction and analysis of a single picture can support the target recognition task; the design concept of ICRN is to use the continuous characteristics of data to realize recognition, so there is no obvious advantage in performance.

[0074] 3) In the case of low SNR, ICRN performs better than the other two methods, and the recognition accuracy is 18% higher than other methods in the observation condition of PD=0. Experimental results show that: when the SNR is low and the information of a single picture is seriously disturbed, ICRN can extract and use the continuous features of the target to realize more efficient recognition, which is in line with the design concept of ICRN.

[0075] In summary, ICRN can achieve the highest recognition effect of about 18% higher than other methods in the case of low SNR, low target PD, and stable distribution of data.

[0076] Further, in step S400, the model generalization and dynamic sequential target recognition performance, the implementation process includes:

[0077] For model generalization, arrange the data according to the observation condition, select a part of the observation condition data for model training, use the remaining part to verify, and analyze the generalization recognition performance;

[0078] For the dynamic sequential target recognition performance of the model, control the offset distance, offset angle, observation azimuth angle and signal-to-noise ratio of the data, and record the dynamic sequential UWB-IR vehicle target recognition performance of the model as the detection distance approaches.

[0079] Specifically, in step S400, the UWB-IR imaging data is divided into two parts according to SA, the data with SA of 90°-270° is used for model training, and the remaining data is used for verification, and the results are shown in Table 6.

[0080] Table 6 Angle generalization test effect

[0081]

[0082] From the above table, it can be found that the basic performance of each method is similar to the previous section, and it is decreased with the increase of PD. In addition, compared with the data of the previous section, it can be found that the recognition effect of CNN decreases obviously with the change of data distribution, which decreases by nearly 10% on average, which is much larger than the 4% and 5% of GoogleNet and ICRN. The analysis reason may be that the rich convolution module of GoogleNet provides sufficient generalization for the model, and ICRN relies on the continuous characteristics of the target, which also ensures the recognition effect when SA changes greatly.

[0083] In addition, in order to more reasonably compare the effect of the sequential target recognition method proposed in this paper, the patent also proposes a comparison test criterion, and compares the effects of ICRN and GoogleNet under the criterion.

[0084] Specifically, first, all imaging results from far to near under fixed observation conditions are randomly selected (the distance is from 10m to 90m); then, for the ICRN model, the number of input images is gradually increased from 0, and when the number of images reaches 30, the data is combined by sliding window backward from 30, so 81 groups of data can be obtained, and 81 experimental results can be obtained; for GoogleNet, each image is identified, and finally 81 identification results can be obtained. For all identification results, output 1 for correct and 0 for error, and count the data under multiple routes, and take the average under each distance, and arrange the identification accuracy of the two models according to the receiving distance of the image. Figure 6 The average effect of the full route recognition results of the two methods under the offset distance of 0m is shown.

[0085] From Figure 6 It can be found that the recognition effect of GoogleNet has little to do with the imaging distance of the image, and it is random; compared with ICRN, the recognition effect of ICRN gradually increases with the decrease of the collection distance of the image. The analysis result shows that the effect of ICRN will gradually increase as the data volume increases from 1 to 30, the effectiveness of the data will increase as the distance of the input imaging data decreases, and the recognition effect of ICRN will also increase significantly. According to the above results and analysis deduction, because ICRN can extract the continuous characteristics of the target image, ICRN can perform better when the data collection distance is short, so it can be proved that ICRN can extract the continuous characteristics of the target image.

[0086] In summary, the embodiment proposes a sequential UWB-IR image vehicle target recognition method based on ICRN, which combines iterative convolution and LSTM, fully extracts and utilizes the continuous change characteristics of the target, and can realize efficient recognition of UWB-IR targets under low signal-to-noise ratio, and provides a new technical approach for UWB-IR target recognition in complex background.

[0087] The above description is only preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A sequential UWB-IR image vehicle target recognition method based on ICRN, characterized in that, Includes the following steps: Step S100: Construction of UWB-IR vehicle target imaging dataset based on array imaging method: The array imaging algorithm is introduced into UWB-IR, the UWB-IR target imaging method is given, and the UWB-IR images of vehicle targets under different observation conditions are simulated according to the given vehicle scattering point model, and the UWB-IR vehicle dataset is constructed. Step S200: ICRN network model design: Design the ICRN model, including the Masking module, the iterative convolution module, and the RNN module. The ICRN model utilizes the depth features of a single image target and the depth correlation features between targets in a sequence of images, while also having dynamic adaptability to the number of sequential images. In step S200, each module specifically includes: The Masking module estimates the dhash calculation between the input image and the reference image, sets the dhash threshold for sequential image selection, and performs image sequence filtering and encoding. The iterable convolutional module includes four convolutional layers with kernel sizes of 5*5, 3*3, 3*3, and 3*3, and each layer has one input and one output channel. Each convolutional layer includes a 3*3 pooling layer with a stride of 2 and a batch normalization layer. In addition to the basic four convolutional layers, it also includes a set of update weights for updating subsequent images. The parameter dimensions of these weights are consistent with those of the basic four-layer convolutional model, i.e., ω. k+1 =ω k ·ω iterable +b iterable , where ω iterable With b iterable To update the weights, ω k ω is the convolution parameter of the previous image. k+1 These are the convolution parameters for the next image; The RNN module includes a 64-node LSTM layer and two fully connected layers, with inputs and outputs of 64*64 and 64*3 respectively. Except for the last fully connected layer, which does not contain an activation function, all other layers that do contain an activation function use ReLU, and the model loss function is the cross-entropy function. Step S300: ICRN Model Performance Validation: Validate the performance of the ICRN model on the simulated UWB-IR vehicle dataset and compare it with other deep network methods; Step S400: Model generalization and dynamic sequential target recognition performance: Use different distribution data of the same target to train and validate ICRN, and analyze the generalization of ICRN based on the performance of the validation set; at the same time, analyze the dynamic sequential UWB-IR vehicle target recognition performance of the model when there is relative motion between the radar and the target.

2. The method for vehicle target recognition based on ICRN sequential UWB-IR images according to claim 1, characterized in that, In step S100, the UWB-IR vehicle target imaging method introduces array imaging into UWB-IR, and provides a UWB-IR target imaging method; based on the given vehicle scattering point model, different observation conditions are simulated, including the target offset distance relative to the radar, offset angle, observation azimuth angle, and UWB-IR images of vehicle targets under the influence of noise, and a UWB-IR vehicle dataset is constructed.

3. The method for vehicle target recognition based on ICRN sequential UWB-IR images according to claim 1, characterized in that, For the Masking module, its input data consists of k consecutive data points in the time dimension, labeled t1 to t2. k The Masking module filters and selects the first m sequential images after filtering, or fills the m images with zeros, and passes them to the iterative convolution module, and marks the number of valid sequence images. For the iterable convolution module, it performs convolution feature extraction on the continuous time-dimensional data passed by the Masking module, the convolution parameters satisfy the transmission relationship, and concatenates the extracted features into a time series according to the time dimension, and passes it to the RNN module. For the RNN module, the LSTM model is used to further extract features from the input feature sequence. Based on the effective sequence length provided by the Masking module, the model update stops after the iterative output reaches the actual sequence length. Finally, the output features are classified using two fully connected layers, and the classification results are output.

4. The method for vehicle target recognition based on ICRN sequential UWB-IR images according to claim 3, characterized in that, The ICRN model uses the Masking module to ensure the continuity and effectiveness of the input data in the time dimension; the ICRN model uses the iterative convolution module to initially extract continuous features from the continuous data; at the same time, the ICRN model does not increase parameters when processing inputs of arbitrary length; the ICRN model further extracts and utilizes the target continuous features by fusing iterative convolution with the RNN module.

5. The method for vehicle target recognition based on ICRN sequential UWB-IR images according to claim 1, characterized in that, In step S300, the ICRN model is validated. On the simulation data generated in step S100, the ICRN model is used to compare and analyze the experimental results with other deep network methods.

6. The method for vehicle target recognition based on ICRN sequential UWB-IR images according to claim 1, characterized in that, In step S400, the model's generalization ability and dynamic sequential target recognition performance are achieved through the following process: To assess model generalization, the data is arranged according to observation conditions. A portion of the observation condition data is selected for model training, and the remaining portion is used for validation. The generalization recognition performance is then analyzed. To assess the dynamic sequential target recognition performance of the model, the offset distance, offset angle, observation azimuth angle, and signal-to-noise ratio of the control data were kept constant, and the dynamic sequential UWB-IR vehicle target recognition performance of the model was recorded as the detection distance approached.

Images