Through-the-wall radar human body posture detection method based on random forest

Abstract

The invention belongs to the technical field of through-the-wall radar human body posture detection, and particularly relates to a through-the-wall radar human body posture detection method based on a random forest, which comprises the following steps: S1, multi-dimensional space-time signal acquisition: adopting a 10-transmitting 10-receiving MIMO-SFCW radar array, using an electromagnetic simulation technology and a signal propagation theory to optimize antenna arrangement, and in a 4m * 4m room, carrying out multi-dimensional space-time signal acquisition; the radar array is arranged at a proper position, a tested person makes three static postures of standing, sitting and lying at a specific distance from the array and keeps static, background data are synchronously collected, and multi-dimensional information such as the distance, the angle and the Doppler signal of a target is obtained. The optimized MIMO-SFCW radar array adopts a unique antenna layout design, target signals can be more accurately captured in a complex environment, external interference signals can be effectively suppressed, the range resolution can be improved by using the improved pulse compression algorithm, and the high classification capability of the random forest algorithm is matched, so that the high-resolution MIMO-SFCW radar array is obtained. And the recognition accuracy of various complex postures is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of through-wall radar human posture detection, and in particular to a through-wall radar human posture detection method based on random forest. Background Art

[0002] In through-wall sensing technology, the detection, identification, and classification of targets behind buildings have always been a key research focus. Traditional gesture recognition relies primarily on optical cameras and depth-of-field cameras, which capture depth maps from optical images to describe the target's position, outline, and shape. However, this approach is inapplicable when there are obstructions due to walls. Although radar sensors can penetrate walls, current through-wall radar target recognition technology focuses on micro-Doppler analysis, which extracts artificial features from range images or Doppler information for recognition. This results in a low recognition rate. Therefore, we propose a through-wall radar human gesture detection method based on random forests to address this issue. Summary of the Invention

[0003] (1) Technical problems solved

[0004] In view of the shortcomings of the existing technology, the present invention provides a through-wall radar human posture detection method based on random forest, which solves the problems raised in the above background technology.

[0005] (2) Technical solution

[0006] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0007] A method for detecting human posture using through-wall radar based on random forests, comprising the following steps:

[0008] S1: Multi-dimensional spatiotemporal signal acquisition: A 10-transmitter, 10-receiver MIMO-SFCW radar array was used. During the antenna arrangement optimization process, electromagnetic simulation technology and signal propagation theory were applied. Considering the propagation characteristics of radar signals in complex environments and the spatial distribution patterns of targets, calculations and repeated simulations were used to rationally adjust the position, spacing, angle, and polarization of the antennas to achieve efficient acquisition of multi-dimensional target information. In an ordinary 4m×4m open room, the radar array was scientifically placed at the bottom edge of the short side of the room. The subjects were asked to stand, sit, and lie down in three static postures in sequence at a distance of 2 meters from the array. They remained stationary while simultaneously collecting background data to obtain multi-dimensional information such as the target's distance, angle, and Doppler signal.

[0009] For MIMO-SFCW radar, its transmit signal can be expressed as:

[0010]

[0011] in, is the amplitude of the nth pulse, is the pulse repetition period, is the pulse width, is the starting frequency, is the frequency step, N is the total number of pulses, and t is the time.

[0012] After the received signal is mixed and low-pass filtered, the difference frequency signal is obtained. The distance R to the target is related to:

[0013]

[0014] in, is the amplitude of the received signal, c is the speed of light, and by processing the difference frequency signal, the distance information of the target can be obtained.

[0015] S2: Signal Processing: Based on the precise principle of MIMO radar phase coherence, the collected radar signals are subjected to interference elimination processing. In this process, signal processing algorithms and mathematical models are used to calculate phase coherence. Through in-depth analysis of the phase relationship of the received signals of each antenna, external interference signals are accurately identified and removed. These interference signals come from surrounding electronic equipment, fluctuations in the electromagnetic environment, etc.

[0016] Assume that the signal received by the i-th receiving antenna and the j-th transmitting antenna is , the phase coherence can be analyzed by calculating the cross-correlation function:

[0017]

[0018] in, represents the mathematical expectation, is the signal received by another transmitting and receiving antenna pair, and τ is the time delay. By analyzing the peak and phase information of the cross-correlation function, the coherence between the signals can be determined, thereby identifying the interference signal.

[0019] At the same time, the frequency domain signal is pulse compressed and the advanced matched filter algorithm and adaptive signal processing technology are used to improve the distance resolution and the impulse response of the matched filter. With the transmission signal Conjugate inversion:

[0020]

[0021] Where T is the signal duration. After matched filtering, the output signal y(t) is:

[0022]

[0023] in, Represents the convolution operation, which can compress the pulse width and improve the distance resolution.

[0024] Combined with the spatial array placement, three-dimensional modeling and coordinate transformation technology are used to generate a 3D radar reflection map in space. This 3D radar reflection map provides an intuitive and accurate data representation for subsequent data processing and posture recognition, making the spatial position and posture characteristics of the target more clearly presented.

[0025] S3: Data preprocessing: The generated 3D radar reflection map is strictly calibrated, and label assignment rules are formulated. The standing posture is assigned label 0, the sitting posture is assigned label 1, the lying posture is assigned label 2, and the blank background is assigned label 3. During the data calibration process, multiple people are used to check and verify multiple times to ensure a one-to-one correspondence between labels and postures. This data calibration method provides a high-quality and accurate data foundation for model training, enabling the machine learning model to more accurately learn the data features of different postures, thereby improving the model training effect and the accuracy of posture recognition.

[0026] S4: Model training: Use SVM (support vector machine), decision tree and random forest algorithms to conduct comprehensive training on the preprocessed data. Use the 5-fold cross-validation method to scientifically divide the data set into 5 subsets. Each time, 4 subsets are selected for training and 1 subset is used for testing. This is repeated 5 times. This cross-validation method can make full use of the information of the data set, effectively evaluate the performance of the model, and avoid overfitting and underfitting problems.

[0027] During the training process of the random forest algorithm, each decision tree randomly selects samples and features for learning and constructs independent classification rules. Assume that the dataset is:

[0028]

[0029] in, is the eigenvector, is the corresponding label. When constructing the kth decision tree, n samples are randomly selected from D with replacement to form a training subset. ,At the same time, d features (d < total number of features) are randomly selected for node splitting.

[0030] The node splitting of the decision tree usually uses criteria such as information gain and Gini index. According to information gain, it is assumed that feature A has V values. , the dataset D is divided into V subsets , then the information gain of feature A on data set D for:

[0031]

[0032] Where H(D) is the information entropy of the data set D:

[0033]

[0034] is the proportion of the k-th class samples in D, K is the number of categories, and the decision tree is constructed by continuously selecting the features with the largest information gain to split the nodes.

[0035] Through random sampling and feature selection strategies, each decision tree can learn data features from different perspectives, enhancing the diversity and generalization ability of the model. This randomized training method can effectively avoid the model from overfitting the training data, improving the reliability and stability of the model in practical applications. Even when the amount of data is limited or the data distribution changes, the model can still maintain good performance.

[0036] S5: Posture Recognition: The newly collected and processed data is input into the trained random forest-based classification model. The random forest model integrates the prediction results of multiple decision trees through a voting mechanism to obtain the final posture classification. Assume that there are M decision trees in the random forest. For a new sample x, the prediction result of the mth decision tree is , then the final classification result y is:

[0037]

[0038] in, is an indicator function that takes the value 1 if the condition is true and 0 otherwise.

[0039] Furthermore, in S1, the antenna arrangement of the MIMO-SFCW radar array has been optimized. The optimized antenna arrangement can collect multi-dimensional effective information of the target to the greatest extent, ensuring the comprehensiveness and accuracy of the data. The optimization principle is to reasonably adjust the position, spacing and angle of the antenna according to the electromagnetic wave propagation characteristics and the spatial distribution law of the target, reduce signal obstruction and interference, and enhance the signal perception capability of targets with different postures.

[0040] Furthermore, in S2, the interference elimination process is based on the precise calculation of the MIMO radar phase coherence, which can effectively identify and remove external interference signals. By establishing a precise phase model, the phase relationship of the received signals of each antenna is analyzed, the target signal and the interference signal are accurately distinguished, and the interference component is removed by using a filtering algorithm. At the same time, the frequency domain signal is pulse compressed. The pulse compression adopts advanced signal processing algorithms and matched filtering algorithms to improve the distance resolution. Combined with the spatial array placement position, a 3D radar reflection map in space is generated, providing intuitive and detailed data for subsequent posture analysis.

[0041] Furthermore, the data calibration process in S3 formulates detailed calibration procedures and specifications, is operated by professionals, and is checked and verified multiple times to ensure calibration accuracy.

[0042] Furthermore, in S4, during the training process of the random forest algorithm, each decision tree randomly selects samples and features for learning and constructs independent classification rules. In this way, each decision tree can learn data features from different angles, thereby improving the generalization ability and anti-overfitting ability of the model. Specifically, each time a decision tree is constructed, a certain proportion of samples and features are randomly extracted, so that there are certain differences between the decision trees, so that the results of multiple decision trees can be combined to obtain more accurate classification judgments.

[0043] Furthermore, in S5, the random forest model integrates the prediction results of multiple decision trees through a voting mechanism to obtain the final posture classification. During the voting process, each decision tree gives a posture classification result based on its own judgment, and then counts the votes for each posture classification result. The posture with the most votes is the final classification result. This method ensures the accuracy and reliability of the recognition results and reduces the misjudgment that may occur in a single decision tree.

[0044] (3) Beneficial effects

[0045] Compared with the existing technology, the present invention provides a through-wall radar human posture detection method based on random forest, which has the following beneficial effects:

[0046] The present invention uses an optimized MIMO-SFCW radar array with a unique antenna layout design, which can more accurately capture target signals in complex environments and effectively suppress external interference signals. It uses an improved pulse compression algorithm to improve distance resolution. Combined with the powerful classification capability of the random forest algorithm, the recognition accuracy of various complex postures is significantly improved, and the average classification accuracy is high.

[0047] The present invention constructs multiple independent decision trees, effectively enhancing the model's ability to resist overfitting. Even when faced with large amounts of data and data sets with unknown characteristics, it can accurately learn complex patterns in the data. The system can operate stably in different scenarios and environmental conditions, whether in urban environments with dense buildings or in industrial areas with complex and changeable electromagnetic environments.

[0048] The present invention adopts a series of scientific and reasonable methods and technologies. The interference cancellation technology is based on the principle of MIMO radar phase coherence, which can accurately identify and remove external interference signals, ensuring that the collected radar signal is purer. The pulse compression technology effectively improves the range resolution through the optimization algorithm, making the acquired target information more accurate. The 5-fold cross-validation method divides the data set into multiple parts for training and verification, comprehensively evaluates the model performance, avoids the problems of overfitting and underfitting of the model, and improves the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the overall method of the present invention;

[0050] Figure 2 This is a flowchart of the random forest algorithm of the present invention, illustrating the construction and classification process of the random forest;

[0051] Figure 3 This is a schematic diagram of radar acquisition in the present invention, showing the layout of the radar array and the position of the test subjects;

[0052] Figure 4 This is a comparison chart of the 3D imaging results of the standing radar of the present invention and the depth map of Kinect;

[0053] Figure 5 This is a comparison chart of the 3D imaging result of the sitting posture radar of the present invention and the depth map of Kinect;

[0054] Figure 6 This is a comparison chart of the lying-down radar 3D imaging result and the Kinect depth map;

[0055] Figure 7 The confusion matrix comparison diagram of the classification effect of the SVM algorithm of the present invention on different postures;

[0056] Figure 8 The confusion matrix comparison diagram of the classification effect of the decision tree algorithm of the present invention on different postures;

[0057] Figure 9 This is a comparison chart of the confusion matrix of the classification effect of the random forest algorithm of the present invention on different postures. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] Example

[0060] like Figure 1-9 As shown, an embodiment of the present invention proposes a method for detecting human posture using a through-wall radar based on random forest, comprising the following steps:

[0061] S1: Multi-dimensional spatiotemporal signal acquisition: Select a room with a size of 4m×4m as the experimental site. The room should be relatively open and free of obvious electromagnetic interference sources. Install a 10-transmit 10-receive MIMO-SFCW radar array at the bottom edge of the short side of the room to ensure the stability of the radar array. Use professional calibration equipment to calibrate the radar array to ensure its transmission and reception performance is normal. Figure 3 As shown, a clear mark is set up 2 meters away from the radar array as the test area for the subjects.

[0062] According to the electromagnetic wave propagation characteristics and the spatial distribution of the target, the position, spacing, angle and polarization mode of the antenna are adjusted; through multiple simulations and actual tests, the optimized antenna arrangement parameters are determined so that the antenna arrangement spacing is within a reasonable range that can effectively improve the signal acquisition effect, reduce signal obstruction and interference, and enhance the signal perception capability of targets with different postures.

[0063] Invite subjects of different ages, genders and body shapes to participate in the experiment; each subject takes three static postures in the test area, standing, sitting and lying down, and maintains each posture for a period of time so that the radar can collect data stably. Figure 4 、 Figure 5 and Figure 6 As shown in the figure, while the subjects were making gestures, background data was collected synchronously to obtain multi-dimensional information such as the target's distance, angle, and Doppler signal. During the collection process, the collection frequency and time interval were strictly controlled to ensure the consistency and accuracy of the collected data. A total of 1000 frames of standing, sitting, lying, and background data were collected, totaling 4386 frames of data.

[0064] For MIMO-SFCW radar, its transmit signal can be expressed as:

[0065]

[0066] in, is the amplitude of the nth pulse, is the pulse repetition period, is the pulse width, is the starting frequency, is the frequency step, N is the total number of pulses, and t is the time.

[0067] After the received signal is mixed and low-pass filtered, the difference frequency signal is obtained. The distance R to the target is related to:

[0068]

[0069] in, is the amplitude of the received signal, c is the speed of light, and by processing the difference frequency signal, the distance information of the target can be obtained.

[0070] S2: Signal Processing: Based on the phase coherence principle of MIMO radar, correlation algorithms are used to analyze the phase relationship of the signals received by each antenna to identify and remove external interference signals. Advanced signal processing technology is used to perform pulse compression on the frequency domain signal. Combined with the spatial array placement, 3D modeling and coordinate transformation technology are used to generate a 3D radar reflection map within the space.

[0071] The specific content is to use the filtering algorithm to eliminate interference from the collected radar signals; by accurately analyzing the phase relationship of the signals received by each antenna, the algorithm is used to accurately identify and remove interference signals generated by surrounding electronic equipment, electromagnetic environment fluctuations, etc., effectively suppressing more than 90% of external interference.

[0072] Assume that the signal received by the i-th receiving antenna and the j-th transmitting antenna is , the phase coherence can be analyzed by calculating the cross-correlation function:

[0073]

[0074] in, represents the mathematical expectation, is the signal received by another transmitting and receiving antenna pair, and τ is the time delay. By analyzing the peak and phase information of the cross-correlation function, the coherence between the signals can be determined, thereby identifying the interference signal.

[0075] Perform pulse compression processing on frequency domain signals, use advanced matched filter algorithm and adaptive signal processing technology to improve distance resolution, which increases the distance resolution by 30% compared with traditional methods, and the impulse response of the matched filter With the transmission signal Conjugate inversion:

[0076]

[0077] Where T is the signal duration. After matched filtering, the output signal y(t) is:

[0078]

[0079] in, Represents the convolution operation, which can compress the pulse width and improve the distance resolution.

[0080] Combined with the spatial array placement, three-dimensional modeling and coordinate transformation technology are used to convert the processed signal into a 3D radar reflection map in space. During the generation process, image enhancement algorithms are used to improve the clarity and contrast of the reflection map, providing a data basis for subsequent data processing and posture recognition.

[0081] S3: Data preprocessing: Perform strict data calibration on the generated 3D radar reflection map. Clearly formulate data calibration rules and calibrate according to the rule of assigning label 0 to standing posture, label 1 to sitting posture, label 2 to lying posture, and label 3 to blank background.

[0082] Data calibration is carried out by professionals according to detailed calibration procedures and specifications, and multiple people conduct multiple checks and verifications, such as double-checking. One person calibrates and the other checks. In case of disagreements, they discuss and determine together to ensure that the labels correspond to the postures one by one, and the accuracy of the calibration is guaranteed, providing a high-quality and accurate data foundation for model training.

[0083] S4: Model training: Use SVM (support vector machine), decision tree and random forest algorithms to conduct comprehensive training on the preprocessed data, such as Figure 7 、 Figure 8 and Figure 9 As shown in the figure, the 5-fold cross-validation method is used to scientifically divide the dataset into 5 subsets. Each time, 4 subsets are selected for training and 1 subset is selected for testing. This is repeated 5 times to make full use of the information in the dataset, effectively evaluate the model performance, and avoid overfitting and underfitting problems.

[0084] During the training of the random forest algorithm, Figure 2 As shown, each decision tree randomly extracts samples from the data set with replacement. Each decision tree randomly selects samples and features for learning and constructs independent classification rules. Assume that the data set is:

[0085]

[0086] in, is the eigenvector, is the corresponding label. When constructing the kth decision tree, n samples are randomly selected from D with replacement to form a training subset. ,At the same time, d features (d < total number of features) are randomly selected for node splitting.

[0087] The node splitting of the decision tree usually uses criteria such as information gain and Gini index. According to information gain, it is assumed that feature A has V values. , the dataset D is divided into V subsets , then the information gain of feature A on data set D for:

[0088]

[0089] Where H(D) is the information entropy of the data set D:

[0090]

[0091] is the proportion of the k-th class samples in D, K is the number of categories, and the decision tree is constructed by continuously selecting the features with the largest information gain to split the nodes.

[0092] Through random sampling and feature selection strategies, each decision tree can learn data features from different perspectives, enhancing the diversity and generalization ability of the model. This randomized training method can effectively avoid the model from overfitting the training data, improving the reliability and stability of the model in practical applications. Even when the amount of data is limited or the data distribution changes, the model can still maintain good performance.

[0093] S5: Posture Recognition: The newly collected and processed data is input into the trained random forest-based classification model. The random forest model integrates the prediction results of multiple decision trees through a voting mechanism to obtain the final posture classification. Assume that there are M decision trees in the random forest. For a new sample x, the prediction result of the mth decision tree is , then the final classification result y is:

[0094]

[0095] in, is an indicator function that takes the value 1 if the condition is true and 0 otherwise.

[0096] Each decision tree gives a posture classification result based on its own judgment, and counts the votes for each posture classification result. The posture with the most votes is the final classification result, ensuring the accuracy and reliability of the recognition result.

[0097] The system was tested in different scenarios, such as urban environments with dense buildings and industrial areas with complex and changeable electromagnetic environments. The test results showed that the system's detection accuracy of human posture behind walls in these scenarios was ≥98%, and the average classification accuracy reached 0.98, verifying the high accuracy and stability of the system in different scenarios.

[0098] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for detecting human posture using through-wall radar based on random forest, characterized by: The steps include: S1: Multi-dimensional spatiotemporal signal acquisition: A 10-transmitter, 10-receiver MIMO-SFCW radar array was used. Electromagnetic simulation technology and signal propagation theory were used to optimize the antenna arrangement. The radar array was placed in a suitable position in a 4m×4m room. The subjects were asked to stand, sit, or lie down at a specific distance from the array and remain still. Background data was simultaneously collected to obtain multi-dimensional information such as the target's distance, angle, and Doppler signal. S2: Signal Processing: Based on the phase coherence principle of MIMO radar, correlation algorithms are used to analyze the phase relationship of the signals received by each antenna to identify and remove external interference signals. Advanced signal processing technology is used to perform pulse compression on the frequency domain signal. Combined with the spatial array placement, 3D modeling and coordinate transformation technology are used to generate a 3D radar reflection map within the space. S3: Data preprocessing: The generated 3D radar reflection map is strictly calibrated according to the rule of assigning label 0 to standing posture, label 1 to sitting posture, label 2 to lying posture, and label 3 to blank background. Multiple people are used to check and verify the data to ensure that the labels correspond to the postures. S4: Model training: The preprocessed data is trained using support vector machines (SVMs), decision trees, and random forest algorithms, using a 5-fold cross-validation method. During random forest training, each decision tree randomly selects samples and features for learning and constructs independent classification rules. S5: Posture Recognition: The newly collected and processed data is input into the trained random forest-based classification model. The random forest model integrates the prediction results of multiple decision trees through a voting mechanism to obtain the final posture classification.

2. The method for detecting human posture using through-wall radar based on random forest according to claim 1, characterized in that: When optimizing the antenna arrangement in S1, the position, spacing, angle and polarization of the antenna are adjusted according to the electromagnetic wave propagation characteristics and the target space distribution law.

3. The method for detecting human posture using through-wall radar based on random forest according to claim 1, characterized in that: The interference elimination in S2 is based on precise phase analysis and uses a special filtering algorithm to remove interference components; the pulse compression adopts an advanced matched filtering algorithm to improve the distance resolution.

4. The method for detecting human posture using through-wall radar based on random forest according to claim 1, characterized in that: The data calibration process in S3 has detailed calibration procedures and specifications, which are operated by professionals and repeatedly checked and verified to ensure calibration accuracy.

5. The method for detecting human posture using through-wall radar based on random forest according to claim 1, characterized in that: When the random forest algorithm in S4 is trained, each decision tree randomly extracts samples from the data set with replacement, and randomly selects some features (d < the total number of features) for node splitting to enhance the diversity and generalization ability of the model and avoid overfitting.

6. The method for detecting human posture using through-wall radar based on random forest according to claim 1, characterized in that: The voting mechanism of the random forest model in S5 is that each decision tree gives a posture classification result based on its own judgment, and the votes for each posture classification result are counted. The posture with the most votes is the final classification result, ensuring the accuracy and reliability of the recognition result.

Images