Gesture recognition method based on DenseNet combined with CBAM attention mechanism

Abstract

The gesture recognition method based on DenseNet combined with CBAM attention mechanism relates to a gesture recognition method based on a millimeter wave radar.The purpose of the present application is to solve the problem of poor noise resistance, easy to miss detection and low gesture recognition accuracy of the existing gesture recognition detection method for the millimeter wave radar. The process is as follows: 1: the radar collects gesture target signals, processes the collected gesture target signals to obtain RTM, DTM, ATM and ETM; the RTM, DTM, ATM and ETM are spliced to obtain a spliced atlas; 2: an expanded atlas data set is obtained; 3: a DenseNet121-CBAM network model is established; 4: a trained network model is obtained; 5: a spliced atlas is obtained; the spliced atlas is input into the trained network model to complete gesture recognition of the to-be-detected gesture target signal. The present application is used in the field of gesture recognition.

Description

Technical Field

[0001] This invention relates to a gesture recognition method based on millimeter-wave radar. Background Technology

[0002] Current mainstream gesture recognition methods suffer from information leakage and poor recognition rates. Gesture recognition based on millimeter-wave radar is unaffected by lighting conditions and has received widespread attention in recent years due to its small size, high accuracy, and strong security. Existing gesture recognition detection methods for millimeter-wave radar can be broadly categorized into static target elimination algorithms based on Moving Target Indication (MTI) or Moving Target Detection (MTD) and target detection algorithms based on Constant False Alarm Rate Detection (CFAR). However, these methods suffer from poor noise resistance and are prone to missing detections, resulting in low gesture recognition accuracy. Summary of the Invention

[0003] The purpose of this invention is to solve the problems of poor noise resistance, easy missed detection, and low accuracy of gesture recognition in existing gesture recognition detection methods for millimeter-wave radar. Therefore, this invention proposes a gesture recognition method based on DenseNet combined with the CBAM attention mechanism.

[0004] The specific process of the gesture recognition method based on DenseNet combined with CBAM attention mechanism is as follows:

[0005] Step 1: The radar collects hand gesture target signals, processes the collected hand gesture target signals, and obtains the time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM).

[0006] The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map.

[0007] Step 2: Perform translation, rotation, magnitude transformation, random cropping, and Gaussian blur operations on the stitched map obtained in Step 1 to obtain the expanded map dataset;

[0008] Step 3: Establish the DenseNet121-CBAM network model;

[0009] Step 4: Input the expanded map dataset from Step 3 into the DenseNet121-CBAM network model established in Step 3 for training until convergence, and obtain the trained DenseNet121-CBAM network model.

[0010] Step 5: The radar acquires the signal of the hand gesture target to be tested, processes the signal to obtain the time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM);

[0011] The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map.

[0012] The stitched image is input into the trained DenseNet121-CBAM network model to complete the gesture recognition of the target gesture signal.

[0013] The beneficial effects of this invention are as follows:

[0014] This invention uses millimeter-wave radar to collect gesture data, performs target detection and feature extraction on the dataset, and finally uses neural networks to perform gesture recognition of 12 gestures from a high-precision perspective.

[0015] This invention is based on millimeter-wave radar echo signals, data preprocessing, target detection, feature extraction, and deep learning, which is significantly different from most current recognition algorithms. Nowadays, gesture recognition simply transforms radar data and directly feeds it into a neural network to increase the recognition rate by increasing the complexity of the network. However, this invention performs sufficient data processing before adding radar signals to the neural network, extracting enough gesture features, so that a high gesture recognition rate can be achieved without excessively increasing the complexity of the neural network.

[0016] This invention proposes a novel gesture recognition target detection method (DRDP) based on distance-Doppler maps, and compares its performance curves with advanced CFAR-type target detection methods, showing a significant performance improvement. At the same time, based on deep learning, a dataset augmentation method based on image transformation is proposed by inputting the hybrid feature map into the DenseNet network, and combined with the CBAM attention mechanism to improve model performance, ultimately achieving a high recognition accuracy of 99.03%. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the principle of single-feature map splicing.

[0019] Figure 3 This is a segmentation result diagram of the AIT segmentation method;

[0020] Figure 4a Schematic diagram of CBAM position 1011;

[0021] Figure 4b Schematic diagram of CBAM location 0010;

[0022] Figure 5a This is a diagram of the DenseNet121-CBAM network model structure.

[0023] Figure 5b This is a schematic diagram of the accuracy rate change curve;

[0024] Figure 5c This is a schematic diagram of the loss variation curve;

[0025] Figure 5d This is a schematic diagram of the confusion matrix for the DenseNet scheme. Detailed Implementation

[0026] Specific Implementation Method 1: The specific process of the gesture recognition method based on DenseNet combined with the CBAM attention mechanism in this implementation method is as follows:

[0027] Step 1: The radar collects hand gesture target signals, processes the collected hand gesture target signals, and obtains the time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM).

[0028] The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map.

[0029] The data collected in step one is used as the labeled training network;

[0030] Step 2: Perform translation, rotation, magnitude transformation, random cropping, and Gaussian blur operations on the stitched map obtained in Step 1 to obtain the expanded map dataset;

[0031] The combined spectra from step one are subjected to translation, rotation, amplitude transformation, random cropping, and Gaussian blurring operations. The resulting new sample combinations form an expanded dataset, with transformation effects as follows: Figure 3 As shown in Table 2, the comparison of the augmentation effect of the transformed dataset is presented in the table.

[0032] Table 2 Comparison of Dataset Expansion Methods

[0033]

[0034] Feature Dataset Augmentation: This invention collects 600 samples per class, which is still relatively small for deep networks like DenseNet, making it prone to overfitting and failing to reach optimal performance. Therefore, the training set needs to be expanded to enhance the robustness and generalization of the model. This paper analyzes the possible situations that may occur during gesture acquisition and proposes a dataset augmentation method based on image transformation (AIT).

[0035] Step 3: Establish the DenseNet121-CBAM network model;

[0036] Step 4: Input the expanded map dataset from Step 3 into the DenseNet121-CBAM network model established in Step 3 for training until convergence, and obtain the trained DenseNet121-CBAM network model.

[0037] Step 5: The radar acquires the signal of the hand gesture target to be tested, processes the signal to obtain the time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM);

[0038] The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map.

[0039] The stitched image is input into the trained DenseNet121-CBAM network model to complete the gesture recognition of the target gesture signal.

[0040] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that, in step one, the radar collects hand gesture target signals, processes the collected hand gesture target signals, and obtains a time-range map (RTM), a time-velocity map (DTM), a time-azimuth map (ATM), and a time-elevation map (ETM).

[0041] The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map.

[0042] The specific process is as follows:

[0043] Step 11: The radar acquires hand gesture target signals, and performs range FFT and velocity FFT on the acquired hand gesture target signals in the range dimension and velocity dimension respectively to obtain the range-Doppler image RD;

[0044] Projecting the distance-Doppler image RD onto the vertical axis yields the time-distance map RTM.

[0045] Projecting the distance-Doppler plot (RD) onto the horizontal axis yields the time-velocity plot (DTM).

[0046] The collected gesture target signal is transformed into the time domain by performing a Fourier transform on the signal, resulting in the time-azimuth map (ATM) and the time-elevation map (ETM).

[0047] Feature Extraction: For the extracted gesture target, range FFT and velocity FFT are performed on its response matrix in the range and velocity dimensions respectively, presenting the gesture target on a range-Doppler map to obtain the time-range map. The concept of micro-Doppler is incorporated into the extraction process of the time-velocity map (DTM). The acquired radar echo signal can be directly transformed to the time domain using Fourier transform, with commonly used methods including short-time Fourier transform and Wigner distribution. The time-angle map can be further subdivided into time-azimuth-time map (ATM) and time-elevation-time map (ETM), the specific implementation of which is shown in the figure below. Then, DOA estimation is performed on each target point, and the estimation results are superimposed according to the angular unit and the corresponding energy value. A single feature map can only represent one aspect of the gesture movement, so effective mixing of the maps is necessary.

[0048] Steps 1 and 2: Preprocess the time-velocity map (DTM) and time-distance map (RTM) obtained in Step 1 to obtain preprocessed time-velocity map (DTM) and time-distance map (RTM);

[0049] Step 13: Combine the Time-Azimuth Map (ATM) and Time-Elevation Map (ETM) obtained in Step 11 with the preprocessed Time-Velocity Map (DTM) and Time-Distance Map (RTM) obtained in Step 12 to obtain the combined composite map MFTM.

[0050] Single feature map stitching: This invention uses a vertical stitching method (that is, four images are stitched vertically together, using the same horizontal axis); for example Figure 2 .

[0051] The other steps and parameters are the same as in Specific Implementation Method 1.

[0052] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the Time-Velocity Map (DTM) and Time-Distance Map (RTM) obtained in Step One or Two are preprocessed to obtain preprocessed Time-Velocity Map (DTM) and Time-Distance Map (RTM); the specific process is as follows:

[0053] Step 121: Standardize the time-velocity plot (DTM) to obtain the standardized time-velocity plot (DTM); the expression is:

[0054] Before the graph mixing, the energy values ​​in each single feature graph are different, which makes it difficult to effectively utilize single feature graphs with generally small energy values, and the neural network has difficulty converging. Therefore, this paper performs a standardization operation on the single feature graphs.

[0055]

[0056] Where, d rl Represents the original energy value of pixel rl in the time-velocity graph (DTM). The value represents the energy of pixel rl in the time-velocity graph (DTM) after normalization. R and L represent the total number of rows and columns of pixels in the DTM, respectively. max{DTM} represents the energy of the highest energy point in the DTM, and min{DTM} represents the energy of the lowest energy point in the DTM.

[0057] Step 122: Align the time axis of the standardized time-velocity plot (DTM) to obtain the time-axis aligned DTM; the specific process is as follows:

[0058] The horizontal axis of the standardized time-velocity map (DTM) is 3×numframe. The bilinear interpolation method is used to scale and align the standardized time-velocity map DTM in the time dimension, and the horizontal axis is uniformly scaled to 2×numframe.

[0059] After standardization, various single-feature maps can be mixed. This project uses a vertical stitching method to stitch the feature maps, which meets the consistency requirements for RTM, ATM, and ETM maps. However, for DTM, due to the use of micro-Doppler processing and the introduction of time-frequency analysis, windowing is required for time-frequency analysis of a single frame signal. Therefore, the horizontal length of the map needs to be greater than numframe. In this invention, the horizontal axis dimension of DTM is 3×numframe. To align their horizontal lengths and ensure consistency in the timelines of each feature expression, bilinear interpolation is used to scale and align each single-feature map in the horizontal (time) dimension, uniformly scaling the horizontal length to 2×numframe.

[0060] Steps 1, 2, and 3: Perform effective region cropping on the Time-Distance Map (RTM) obtained in Step 1 and the Time-Velocity Map (DTM) obtained in Step 1, 2, and 2. The specific process is as follows:

[0061] For the time-range map (RTM) obtained in step one, only the target echo information of 0-35 range cells is retained in the vertical axis direction, and the remaining range cells are deleted.

[0062] The time-velocity graph (DTM) obtained after time axis alignment in steps one through two has a total of 256 velocity elements. Only velocity elements from the 30th to the 220th are retained, and the remaining velocity elements are deleted.

[0063] For ATM and ETM, since operators may not be directly facing the radar antenna when performing gestures, the gestures may appear in various angular regions. Therefore, each effective angular interval needs to be preserved. Since the scale of ATM and ETM maps is 1°, their information representation scale is consistent and they can be directly stitched together. For RTM and DTM, the situation is different. Since the effective range of the gesture is preset during the gesture design, the gesture target can only appear in a limited number of range cells, unlike angular information which can appear in the entire measurable area. Therefore, RTM does not need to retain all range cell information in the vertical axis direction. According to the gesture action design requirements, retaining the target echo information of 0-35 range cells is sufficient, which greatly reduces the feature size. Similarly, for DTM, there are a total of 256 velocity units, and the maximum velocity that can be displayed on the graph is close to 2 m / s. Under normal circumstances, the speed of hand gestures is generally not that high. Therefore, velocity units with excessively large absolute velocity values ​​can be deleted when splicing and mixing the graphs. In this project, only the 30th to 220th velocity units were retained, and the remaining velocity units were deleted to eliminate redundant feature dimensions.

[0064] Other steps and parameters are the same as in specific implementation method one or two.

[0065] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that step three involves establishing a DenseNet121-CBAM network model; the specific process is as follows:

[0066] The DenseNet121-CBAM network model consists of the following layers in sequence: first convolutional layer, max pooling layer, Dense Block 1 layer, first transition layer, Dense Block 2 layer, second transition layer, CBAM layer, Dense Block 3 layer, third transition layer, Dense Block 4 layer, fourth average pooling layer, softmax layer, and classification layer.

[0067] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0068] Specific Implementation Method 5: This implementation method differs from Specific Implementation Methods 1 to 4 in that the kernel size of the first convolutional layer is 7×7 and the stride is 2.

[0069] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0070] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the DenseBlock1 layer includes a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a seventh convolutional layer;

[0071] The kernel size of the second, fourth, and sixth convolutional layers is 1×1;

[0072] The kernel size of the third, fifth, and seventh convolutional layers is 3×3.

[0073] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0074] Specific Implementation Method Seven: This implementation method differs from one of Specific Implementation Methods One to Six in that the DenseBlock2 layer includes an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, an eleventh convolutional layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a seventeenth convolutional layer, an eighteenth convolutional layer, and a nineteenth convolutional layer.

[0075] The kernel size of the eighth, tenth, twelfth, fourteenth, sixteenth, and eighteenth convolutional layers is 1×1.

[0076] The kernel size of the ninth, eleventh, thirteenth, fifteenth, seventeenth, and nineteenth convolutional layers is 3×3.

[0077] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0078] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that the DenseBlock3 layer includes the twentieth convolutional layer, the twenty-first convolutional layer, the twenty-second convolutional layer, the twenty-third convolutional layer, the twenty-fourth convolutional layer, the twenty-fifth convolutional layer, the twenty-sixth convolutional layer, the twenty-seventh convolutional layer, the twenty-eighth convolutional layer, the twenty-ninth convolutional layer, the thirtieth convolutional layer, the thirty-first convolutional layer, the thirty-second convolutional layer, the thirty-third convolutional layer, the thirty-fourth convolutional layer, the thirty-fifth convolutional layer, the thirty-sixth convolutional layer, the thirty-seventh convolutional layer, the thirty-eighth convolutional layer, the thirty-ninth convolutional layer, the fortieth convolutional layer, the forty-first convolutional layer, the forty-second convolutional layer, and the forty-third convolutional layer.

[0079] The kernel size of the 20th, 22nd, 24th, 26th, 28th, 30th, 32nd, 34th, 36th, 38th, 40th, and 42nd convolutional layers is 1×1.

[0080] The kernel size of the 21st, 23rd, 25th, 27th, 29th, 31st, 33rd, 35th, 37th, 39th, 41st, and 43rd convolutional layers is 3×3.

[0081] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0082] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that the DenseBlock4 layer includes the forty-fourth convolutional layer, the forty-fifth convolutional layer, the forty-sixth convolutional layer, the forty-seventh convolutional layer, the forty-eighth convolutional layer, the forty-ninth convolutional layer, the fiftieth convolutional layer, the fifty-first convolutional layer, the fifty-second convolutional layer, the fifty-third convolutional layer, the fifty-fourth convolutional layer, the fifty-fifth convolutional layer, the fifty-sixth convolutional layer, the fifty-seventh convolutional layer, the fifty-eighth convolutional layer, and the fifty-ninth convolutional layer;

[0083] The kernel size of the forty-fourth, forty-sixth, forty-eighth, fiftieth, fifty-second, fifty-fourth, fifty-sixth, and fifty-eighth convolutional layers is 1×1.

[0084] The kernel size of the 45th, 47th, 49th, 51st, 53rd, 55th, 57th, and 59th convolutional layers is 3×3.

[0085] The first transition layer includes a sixtieth convolutional layer and a first average pooling layer;

[0086] The second transition layer includes a sixty-first convolutional layer and a second average pooling layer;

[0087] The third transition layer includes a sixty-second convolutional layer and a third average pooling layer;

[0088] The kernel size of the 60th convolutional layer is 1×1;

[0089] The kernel size of the sixty-first convolutional layer is 1×1;

[0090] The kernel size of the sixty-second convolutional layer is 1×1.

[0091] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0092] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that, in step four, the expanded atlas dataset from step three is input into the DenseNet121-CBAM network model established in step three for training until convergence, resulting in a trained DenseNet121-CBAM network model; the specific process is as follows:

[0093] Step 3: The expanded atlas dataset is input into a convolutional layer. The output of the convolutional layer is input into a max pooling layer. The output of the max pooling layer is input into a Dense Block 1 layer. The output of the Dense Block 1 layer is input into the first transition layer. The output of the first transition layer is input into a Dense Block 2 layer. The output of the Dense Block 2 layer is input into the second transition layer. The output of the second transition layer is input into a CBAM layer. The output of the CBAM layer is input into a Dense Block 3 layer. The output of the Dense Block 3 layer is input into the third transition layer. The output of the third transition layer is input into a Dense Block 4 layer. The output of the Dense Block 4 layer is input into a fourth average pooling layer. The output of the fourth average pooling layer is input into a softmax layer. The output of the softmax layer is input into a classification layer. The classification layer outputs the classification result.

[0094] Continue until convergence, and you will obtain the trained DenseNet121-CBAM network model.

[0095] The size of the map data after the expansion in step three is 1×320×56.

[0096] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0097] DenseNet's basic gesture recognition model: In an L-layer network model, a typical network can only generate L connections, while DenseNet generates L(L+1) / 2 connections. These dense connections are implemented in dense blocks, each block consisting of several dense layers. This structure ensures that features are not lost during transmission. To reduce the number of parameters, a design parameter Reduction is added to the Transition layer, limiting the number of output parameters to half the number of input parameters. Thanks to this design, the number of network parameters is significantly reduced. Experiments were conducted using five feature representations: extracted time-azimuth spectrum (ATM), time-elevation spectrum (ETM), time-range spectrum (RTM), time-velocity spectrum (DTM), and a hybrid feature map (MFTM) composed of four single feature maps. The experimental accuracy results for each network model are shown in Table 1.

[0098] Table 1. Comparison of feature extraction accuracy between the MFTM model and other models.

[0099]

[0100]

[0101] DenseNet combined with CBAM attention mechanism for classification: Generally, some parts considered less important for acquiring key information are appropriately ignored, which can improve the efficiency and accuracy of people's understanding and perception of their surroundings. This is the starting point of the attention mechanism. To improve model performance, this invention introduces the CBAM attention mechanism based on DenseNet. The CBAM attention mechanism performs attention calculation at both the channel and spatial levels, consisting of two sub-modules: a channel attention module and a spatial attention module. Its processing flow is as follows: (1) Channel attention module. (2) Spatial attention module.

[0102] DenseNet121 contains four Dense Blocks, i.e., four convolutional blocks. To find the optimal way to combine CBAM with DenseNet121, CBAMs were introduced at different positions. Gesture classification was then performed based on the MFTM feature maps augmented using the AIT method. The insertion of the CBAM module was represented by a 4-bit binary code, for example, 0011 represents inserting CBAM after Dense Block 1 and Dense Block 2, and so on. Different CBAM positions had different impacts on the network model performance; at position 1011, even a negative optimization phenomenon occurred. At position 0010, the performance was nearly 1% better than the original DenseNet121 network. To analyze the reasons for this phenomenon, Grad-Cam was used to plot the time-azimuth attention distribution heatmaps of the model for a certain input mixed feature map at positions 1011 and 0010, as shown below. Figure 4a and Figure 4b As shown in Table 3, the data amplification effects generated at the CBAM locations are compared.

[0103] Table 3 Comparison of the effects of different locations of CBAM on dataset augmentation

[0104]

[0105] The proposed new network structure diagram can be referenced. Figure 5a for:

[0106] Input feature map → Max pooling layer → Channel concatenation → 7x7 convolution → Sigmoid → Weight coefficients

[0107] The beneficial effects of the present invention are verified using the following embodiments:

[0108] In summary, this invention adopts the form of adding CBAM after Dense Block2, and the overall network structure after combination is as follows: Figure 5a As shown. The MFTM feature map after data augmentation using the above network structure combined with the AIT method is used as the final high-precision gesture classification scheme in this paper (hereinafter referred to as the DenseNet scheme). The changes in classification accuracy and loss value during the model training process are shown in the figure. Figure 5b , Figure 5c As shown in the figure, during the iteration process, the network is gradually optimized until convergence, and the overall accuracy on the test set steadily increases and remains stable at a certain level. The final overall classification accuracy is 99.03%, and its confusion matrix is ​​shown in the figure. Figure 5dAs shown, after a series of optimization steps including feature map mixing, feature dataset augmentation, and the addition of the CBAM attention mechanism, the final recognition rate can reach over 98%. For some micro-gestures with small amplitude and no hand movement, the system adapts well to micro-gestures due to the addition of micro-Doppler information in the MFTM map, achieving a recognition accuracy of over 96%. Overall, the system can achieve high-accuracy classification of the 12 collected gesture actions.

[0109] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A gesture recognition method based on DenseNet combined with CBAM attention mechanism, characterized in that: The specific process of the method is as follows: Step 1: The radar collects hand gesture target signals, processes the collected hand gesture target signals, and obtains the time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM). The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map. Step 2: Perform translation, rotation, magnitude transformation, random cropping, and Gaussian blur operations on the stitched map obtained in Step 1 to obtain the expanded map dataset; Step 3: Establish the DenseNet121-CBAM network model; the specific process is as follows: The DenseNet121-CBAM network model consists of the following layers in sequence: first convolutional layer, max pooling layer, Dense Block 1 layer, first transition layer, Dense Block 2 layer, second transition layer, CBAM layer, Dense Block 3 layer, third transition layer, Dense Block 4 layer, fourth average pooling layer, softmax layer, and classification layer. The kernel size of the first convolutional layer is The step size is 2; The Dense Block1 layer includes a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, and a seventh convolutional layer; The kernel sizes of the second, fourth, and sixth convolutional layers are: ; The kernel sizes of the third, fifth, and seventh convolutional layers are: ; The Dense Block2 layer includes the eighth, ninth, tenth, eleventh, twelfth, thirteenth, fourteenth, fifteenth, sixteenth, seventeenth, eighteenth, and nineteenth convolutional layers. The kernel sizes of the eighth, tenth, twelfth, fourteenth, sixteenth, and eighteenth convolutional layers are: ; The kernel sizes of the ninth, eleventh, thirteenth, fifteenth, seventeenth, and nineteenth convolutional layers are: ; The Dense Block3 layer includes the twentieth convolutional layer, the twenty-first convolutional layer, the twenty-second convolutional layer, the twenty-third convolutional layer, the twenty-fourth convolutional layer, the twenty-fifth convolutional layer, the twenty-sixth convolutional layer, the twenty-seventh convolutional layer, the twenty-eighth convolutional layer, the twenty-ninth convolutional layer, the thirtieth convolutional layer, the thirty-first convolutional layer, the thirty-second convolutional layer, the thirty-third convolutional layer, the thirty-fourth convolutional layer, the thirty-fifth convolutional layer, the thirty-sixth convolutional layer, the thirty-seventh convolutional layer, the thirty-eighth convolutional layer, the thirty-ninth convolutional layer, the fortieth convolutional layer, the forty-first convolutional layer, the forty-second convolutional layer, and the forty-third convolutional layer. The kernel sizes of the 20th, 22nd, 24th, 26th, 28th, 30th, 32nd, 34th, 36th, 38th, 40th, and 42nd convolutional layers are... ; The kernel sizes of the 21st, 23rd, 25th, 27th, 29th, 31st, 33rd, 35th, 37th, 39th, 41st, and 43rd convolutional layers are... ; The Dense Block4 layer includes the 44th, 45th, 46th, 47th, 48th, 49th, 50th, 51st, 52nd, 53rd, 54th, 55th, 56th, 57th, 58th, and 59th convolutional layers. The kernel sizes of the 44th, 46th, 48th, 50th, 52nd, 54th, 56th, and 58th convolutional layers are: ; The kernel sizes of the 45th, 47th, 49th, 51st, 53rd, 55th, 57th, and 59th convolutional layers are: ; The first transition layer includes a sixtieth convolutional layer and a first average pooling layer; The second transition layer includes a sixty-first convolutional layer and a second average pooling layer; The third transition layer includes a sixty-second convolutional layer and a third average pooling layer; The kernel size of the sixtieth convolutional layer is... ; The kernel size of the sixty-first convolutional layer is ; The kernel size of the sixty-second convolutional layer is ; Step 4: Input the expanded map dataset from Step 3 into the DenseNet121-CBAM network model established in Step 3 for training until convergence, and obtain the trained DenseNet121-CBAM network model. Step 5: The radar acquires the signal of the hand gesture target to be tested, processes the signal to obtain the time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM); The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map. The stitched image is input into the trained DenseNet121-CBAM network model to complete the gesture recognition of the target gesture signal.

2. The gesture recognition method based on DenseNet combined with CBAM attention mechanism according to claim 1, characterized in that: In step one, the radar collects hand gesture target signals, processes the collected hand gesture target signals, and obtains time-range map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM). The obtained time-distance map (RTM), time-velocity map (DTM), time-azimuth map (ATM), and time-elevation map (ETM) are stitched together to obtain the stitched map. The specific process is as follows: Step 11: The radar acquires hand gesture target signals, and performs range FFT and velocity FFT on the acquired hand gesture target signals in the range dimension and velocity dimension respectively to obtain the range-Doppler image RD; Projecting the distance-Doppler image RD onto the vertical axis yields the time-distance map RTM. Projecting the distance-Doppler plot (RD) onto the horizontal axis yields the time-velocity plot (DTM). The collected gesture target signal is transformed into the time domain by performing a Fourier transform on the signal, resulting in the time-azimuth map (ATM) and the time-elevation map (ETM). Steps 1 and 2: Preprocess the time-velocity map (DTM) and time-distance map (RTM) obtained in Step 1 to obtain preprocessed time-velocity map (DTM) and time-distance map (RTM); Step 13: Combine the Time-Azimuth Map (ATM) and Time-Elevation Map (ETM) obtained in Step 11 with the preprocessed Time-Velocity Map (DTM) and Time-Distance Map (RTM) obtained in Step 12 to obtain the combined composite map MFTM.

3. The gesture recognition method based on DenseNet combined with CBAM attention mechanism according to claim 2, characterized in that: In steps one and two, the time-velocity map (DTM) and time-distance map (RTM) obtained in step one are preprocessed to obtain preprocessed time-velocity map (DTM) and time-distance map (RTM); the specific process is as follows: Step 121: Standardize the time-velocity plot (DTM) to obtain the standardized time-velocity plot (DTM); the expression is: in, Represents pixels in the time-velocity graph (DTM) The original energy value, Represents pixels in the time-velocity graph (DTM) Energy value after performing standardized operations. and These represent the total number of rows and columns of pixels in the time-velocity DTM, respectively. This represents the energy of the highest energy point in the DTM graph. This represents the energy of the lowest energy point in the DTM graph; Step 122: Align the time axis of the standardized time-velocity plot (DTM) to obtain the time-axis aligned DTM; the specific process is as follows: The horizontal axis dimension of the standardized time-velocity plot DTM is: The standardized time-velocity map (DTM) was scaled and aligned in the time dimension using bilinear interpolation, with the horizontal axis dimension uniformly scaled to [size missing]. size; Steps 1, 2, and 3: Perform effective region cropping on the Time-Distance Map (RTM) obtained in Step 1 and the Time-Velocity Map (DTM) obtained in Step 1, 2, and 2. The specific process is as follows: For the time-range map (RTM) obtained in step one, only the target echo information of 0-35 range cells is retained in the vertical axis direction, and the remaining range cells are deleted. The time-velocity graph (DTM) obtained after time axis alignment in steps one through two has a total of 256 velocity elements. Only velocity elements from the 30th to the 220th are retained, and the remaining velocity elements are deleted.

4. The gesture recognition method based on DenseNet combined with CBAM attention mechanism according to claim 3, characterized in that: In step four, the expanded atlas dataset from step three is input into the DenseNet121-CBAM network model established in step three for training until convergence, resulting in a trained DenseNet121-CBAM network model. The specific process is as follows: Step 3: The expanded atlas dataset is input into the first convolutional layer. The output of the convolutional layer is input into the max pooling layer. The output of the max pooling layer is input into the Dense Block 1 layer. The output of the Dense Block 1 layer is input into the first transition layer. The output of the first transition layer is input into the Dense Block 2 layer. The output of the Dense Block 2 layer is input into the second transition layer. The output of the second transition layer is input into the CBAM layer. The output of the CBAM layer is input into the Dense Block 3 layer. The output of the Dense Block 3 layer is input into the third transition layer. The output of the third transition layer is input into the Dense Block 4 layer. The output of the Dense Block 4 layer is input into the fourth average pooling layer. The output of the fourth average pooling layer is input into the softmax layer. The output of the softmax layer is input into the classification layer. The classification layer outputs the classification result. Continue until convergence, and you will obtain the trained DenseNet121-CBAM network model. The size of the map data in the expanded map dataset after step three is: .

Images