Public safety area gait recognition method based on Lora

Abstract

The invention relates to the field of public safety monitoring, in particular to a public safety area gait recognition method based on Lora, which comprises the following steps of: 1, deleting null values in LoRa signals, converting the LoRa signals in a complex number form into time sequence signals in a real number field, processing the signals by using a differential signal processing method, and highlighting signal changes caused by gaits; 2, the differential signals are remodeled and converted into an image format to meet the model training requirement, the converted images and corresponding labels are input into a convolutional neural network for training, and effective classification of gait signals is achieved; and step 3, a self-adaptive semi-supervised deep clustering algorithm: the convolutional neural network is trained by adopting the self-adaptive semi-supervised deep clustering algorithm, and the trained convolutional neural network is used for gait recognition. According to the method, the LoRa signal can be ingeniously utilized to collect and analyze the gait features, so that the recognition accuracy is remarkably improved, and the anti-interference capability is greatly enhanced.

Description

Technical Field

[0001] The present invention relates to the field of public security monitoring, and particularly to the field of LoRa sensing technology. Specifically, it is a gait recognition method for public security areas based on LoRa. Background Art

[0002] Traditional public security monitoring systems mainly rely on video surveillance, combined with manual observation and recognition to ensure security. However, this mode has significant drawbacks: it not only requires a large amount of human input, but also the long-term duty of monitoring personnel is prone to fatigue, which in turn weakens the actual monitoring effect. In addition, video surveillance technology itself is also easily restricted by various factors such as light changes, obstacles, and bad weather, resulting in limited recognition accuracy and frequent misjudgments. More importantly, video surveillance is unable to effectively track and identify personnel in the long term, and it is difficult to effectively handle emergencies such as missing persons and escapes. With the rapid development of technology, gait recognition, as a cutting-edge biometric technology, is showing unprecedented application potential in the field of public security. This technology realizes long-distance, non-contact, and relatively stable personnel identity recognition and behavior analysis by carefully analyzing the gait characteristics of individuals when walking, including subtle differences such as walking speed, stride length, and walking frequency. Summary of the Invention

[0003] In order to solve the problems of the traditional public security monitoring system, which requires a large amount of human input and has a poor actual monitoring effect, the present invention proposes a gait recognition method for public security areas based on LoRa.

[0004] The present invention is implemented by the following technical scheme: A gait recognition method for public security areas based on LoRa, which is realized through the following steps:

[0005] Step 1: Gait signal collection and preprocessing: Delete the null values in the original LoRa signal, convert the complex LoRa signal into a time series signal in the real number domain, and then process the signal using the differential signal processing method to highlight the signal changes caused by gait.

[0006] Step 2: Gait feature extraction: Reshape and convert the differential signal into an image format to meet the model training requirements, and input the converted image and the corresponding label into a convolutional neural network for training to achieve effective classification of gait signals.

[0007] Step 3: Adaptive semi-supervised deep clustering algorithm: The convolutional neural network is trained using the adaptive semi-supervised deep clustering algorithm, and the trained convolutional neural network is used for gait recognition.

[0008] For the above-mentioned gait recognition method for public security areas based on LoRa, the specific process of Step 1 is as follows:

[0009] Step 1.1: Simultaneously collect the LoRa signals of each of the two antennas. The LoRa signal collected by one antenna is the active LoRa signal, and the LoRa signal collected by the other antenna is the inactive LoRa signal. The LoRa signal is represented as a complex number, and read the collected LoRa signals.

[0010] Step 1.2: After reading, delete the null values in the LoRa signals.

[0011] Step 1.3: Convert the complex number to the real number domain by extracting the real part, thereby generating a continuous time series signal sequence.

[0012] Step 1.4: Adopt differential signal processing, subtract the time series signal sequence formed by the inactive LoRa signal from the time series signal sequence formed by the active LoRa signal to form a differential signal.

[0013] For the above LoRa-based gait recognition method in public safety areas, the specific process of step 2 is as follows:

[0014] Step 2.1: Extract key features: Reshape the differential signal, flatten the differential signal into a one-dimensional linear array, and perform padding or truncation. Normalize the reshaped data and scale it to a standard numerical range. Convert the normalized data into an image format.

[0015] Step 2.2: Assign corresponding labels to each image data sample: Let the image data set be data, and its corresponding label set be labels. If labels[i] is equal to 1, it indicates that the label of the i-th image data sample is 0. Through such settings, the data set is divided into a labeled data set truedata and an unlabeled data set Nolabeldata. After the preparation work is completed, input these two data sets into the convolutional neural network for training so that the model can learn and recognize the features of different category data, thereby achieving effective classification.

[0016] For the above LoRa-based gait recognition method in public safety areas, the specific process of step 3 is as follows:

[0017] Step 3.1: Input the labeled and unlabeled data sets into the convolutional neural network for training.

[0018] Step 3.2: Select data points in the unlabeled data set as the initial clustering centers. For each data point in the data set, calculate its distance from each clustering center and assign it to the nearest clustering center.

[0019] Step 3.3: Recalculate the cluster center for each cluster. The new cluster center is the mean of all data points within the cluster, and the calculation formula is as follows: Where: S i is the data set of the i-th cluster center, and |S i | is the number of data points in this set. This step is to repeat the assignment and update steps until the termination condition is met, that is, the distance between the new cluster center and the old cluster center is less than the preset threshold; at this time, each unlabeled data has a possible confidence class around a cluster center. By calculating the similarity between each unlabeled data and the corresponding cluster center and normalizing it to the range of [0, 1], data with different confidence levels are obtained;

[0020] Step 3.4: In each iteration, select an optimal threshold, which divides the non-confidence values of the data into two groups with the largest between-class variance. After traversing all possible thresholds, a segmentation threshold that can maximize the between-class variance is determined; subsequently, for each class, if the number of samples with a confidence level lower than the threshold exceeds 10, the same method is used to continue searching for an updated threshold. Finally, pseudo-labeled samples with high confidence levels are screened out for each class and these samples are saved. By setting a confidence threshold, a data set with pseudo-labels is created and put into the model to expand the data set; specifically, unlabeled data with a confidence level higher than the set threshold is regarded as pseudo-labeled data, and these data are input into the CNN together with the labeled data for the next round of training iteration. Through this process, the classification accuracy of the model is continuously improved.

[0021] In the above method for gait recognition in a public safety area based on LoRa, the Euclidean distance is used as the distance metric in step 3.2, and the calculation formula is as follows: Where x is the data, and c i is the i-th cluster center, d is the data dimension, x j , c ij are the values of x and c i on the j-th dimension respectively.

[0022] The present invention can cleverly use LoRa signals to collect and analyze gait features, not only significantly improving the recognition accuracy but also greatly enhancing the anti-interference ability. This breakthrough design will effectively improve the efficiency and accuracy of public safety monitoring, providing strong support for building a safer and more intelligent public environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a system framework diagram of a gait recognition method in a public safety area based on Lora.

[0024] Figure 2Flowchart of an adaptive semi - supervised clustering algorithm for a LoRa - based gait recognition method in public safety areas. Detailed implementation manners

[0025] A LoRa - based gait recognition method in public safety areas, which is realized through the following steps:

[0026] Step 1: Gait signal collection and pre - processing: The collected LoRa signals are often accompanied by noise, which increases the difficulty of extracting gait features. In the face of the problems of phase offset and noise interference that may occur during the acquisition of LoRa signals, the present invention processes the signals by deleting the null values in the original LoRa signals and converting the complex - form LoRa signals into time - series signals in the real - number domain. Then, the differential signal processing method is used to highlight the signal changes caused by gait. The specific steps are as follows:

[0027] Step 1.1: Simultaneously collect the LoRa signals of each of the two antennas. The LoRa signal collected by one antenna is the active LoRa signal (the signal collected during gait activity), and the LoRa signal collected by the other antenna is the inactive LoRa signal (the signal collected when there is no gait activity). LoRa signals are usually represented as complex numbers, and the collected LoRa signals are read.

[0028] Step 1.2: After reading, delete the null values in the LoRa signals.

[0029] Step 1.3: Convert the complex numbers into the real - number domain by extracting the real part, so as to generate a continuous time - series signal sequence.

[0030] Step 1.4: Adopt differential signal processing, subtract the time - series signal sequence formed by the inactive LoRa signal from the time - series signal sequence formed by the active LoRa signal to form a differential signal.

[0031] Differential signal = Active data - Inactive data

[0032] If the lengths of the time - series signal sequences are inconsistent, they can be aligned by truncation or interpolation methods, and for the signals of each antenna, the differential signals are calculated point by point.

[0033] In the differential signal, the signal changes caused by gait will be prominently displayed, while the influence of background noise or static environment will be weakened. Finally, a time - series graph is drawn to observe the changes in the differential signal. Through this method, the obvious signal changes caused by gait can be highlighted.

[0034] Step 2: Gait Feature Extraction: Reshape the differential signal and convert it into an image format of 28x28 pixels to meet the requirements of model training. Input the processed data and corresponding labels into a Convolutional Neural Network (CNN) for training to achieve effective classification of gait signals. The specific steps are as follows:

[0035] Step 2.1: Extract key features from the denoised data. By reshaping the differential signal, flatten the differential signal into a one-dimensional linear array, and appropriately pad or truncate it. Normalize the reshaped data and scale it to a standard numerical range to meet the size requirements of 28x28. Subsequently, convert the standardized data into an image format of 28x28 pixels to improve the efficiency and accuracy of model processing.

[0036] Step 2.2: Assign corresponding labels to each image data sample. Let the image dataset be data, and its corresponding label set be labels. If labels[i] is equal to 1, it indicates that the label of the i-th image data sample is 0. Through such settings, the dataset is divided into a labeled dataset truedata and an unlabeled dataset Nolabeldata. After the preparation work is completed, these two datasets can be input into the Convolutional Neural Network (CNN) for training so that the model can learn and identify the features of different categories of data, thereby achieving effective classification.

[0037] Step 3: Adaptive Semi-Supervised Deep Clustering Algorithm: The adaptive semi-supervised deep clustering algorithm is a technique that combines deep learning and clustering analysis. It can utilize a large amount of unlabeled data to perform effective clustering tasks with only a small amount of labeled data. This algorithm is particularly useful when dealing with high-dimensional data and complex data structures. The method is mainly divided into four parts: training the model, using the trained model to classify unlabeled data, semi-supervised clustering, and selecting pseudo-labeled data with high confidence to participate in the next round of model training.

[0038] Step 3.1: Input the labeled and unlabeled datasets into the Convolutional Neural Network (CNN) for training. The main method is as follows:

[0039] Input:

[0040] V = conv2(W, X, "valid") + b

[0041] Output:

[0042] Y = φ(V)

[0043] The input-output formula above is for each convolutional layer. Each convolutional layer has a different weight matrix W, and W, X, and Y are in matrix form. For the last fully connected layer, denoted as the L-th layer, the output is in vector form y L , with the expected output being d, then there is the total error formula:

[0044]

[0045] conv2() represents the convolution operation. The parameter valid specifies the type of convolution operation. W is the convolution kernel matrix, X is the input matrix, which is the labeled dataset truedata and the unlabeled dataset Nolabeldata. b is the bias, φ(x) is the activation function. In the total error, d and y are the vectors of the expected output and the network output respectively, and ||x||2 represents the 2-norm of the vector x

[0046] The gradient formulas for the convolutional layer and the pooling layer are:

[0047]

[0048] When the number of convolution kernels is N

[0049]

[0050] Finally, we get:

[0051]

[0052] Step 3.2: Select appropriate data points in the unlabeled dataset as the initial cluster centers. Assign each data point to the nearest cluster center. For each data point in the dataset, calculate its distance from each cluster center and assign it to the nearest cluster center. Euclidean distance is usually used as the distance metric in this step, and the calculation formula is as follows:

[0053]

[0054] where x is the data point, c i is the i-th cluster center, d is the data dimension, x j , c ij are the values of x and c i on the j-th dimension respectively

[0055] Step 3.3: Recalculate the center of each cluster. For each cluster, recalculate its cluster center. The new cluster center is the mean of all data points within the cluster, and the calculation formula is as follows:

[0056]

[0057] where: Si is the set of data points of the i-th clustering center, |S i | is the number of data points in this set. This step repeats the assignment and update steps until the termination condition is met, that is, the distance between the new clustering center and the old clustering center is less than the preset threshold. By using the K-means clustering algorithm, at this time, each unlabeled data has possible confidence classes around a cluster value. By calculating the similarity between each unlabeled data and the corresponding center cluster value and normalizing it to the range of [0, 1], data with different confidence levels can be obtained. The K-means algorithm is used to assign data points to the closest cluster center and calculate its similarity to the cluster center.

[0058] Step 3.4: Preset threshold. Inspired by the maximum between-class variance method, a method of adaptively iteratively determining the optimal threshold of non-confidence is used to select unlabeled data with high confidence. In each iteration, an optimal threshold is selected, which can divide the non-confidence values of the data into two groups with the maximum between-class variance. The sample of non-confidence values is calculated by the following formula. The lower the non-confidence value, the higher the confidence level.

[0059]

[0060] Among them, in step 3.2, the trained model is used to classify unlabeled training data. For unlabeled training data, its probability value is which reflects its possible class membership. Here, w and b are the parameters of the model, and h(x) represents the function of the model. In step 3.3, based on the features extracted from the deep learning model, constrained seeded k-means clustering is performed on both the labeled dataset and the unlabeled dataset, with the data with clear labels as the constraint conditions. After clustering, the similarity between each unlabeled data and the corresponding cluster center is calculated as and it is normalized to the range of [0, 1].

[0061] Use g to represent the between-class variance of the segmentation threshold:

[0062] g = w o × w 1 × (u o - u 1 ) 2

[0063] where w o and w 1 respectively represent the probabilities of the classification accuracy of the data samples being less than and greater than their non-confidence values, u o and u 1It is their group mean. After traversing all possible thresholds, a segmentation threshold that maximizes the between-class variance is determined. Subsequently, for each class, if the number of samples with confidence lower than the threshold exceeds 10, the same method is used to continue searching for an updated threshold. Finally, pseudo-label samples with high confidence are selected for each class and these samples are saved. A dataset with pseudo-labels is created by setting a confidence threshold and this dataset is put into the model to expand the dataset. Specifically, unlabeled data with confidence higher than the set threshold is regarded as pseudo-labeled data and this data is input into the CNN together with the labeled data for the next round of training iteration. Through this process, the classification accuracy of the model can be continuously improved.

[0064] Through repeated iterations until the model converges, self-supervised clustering of the data is achieved. Based on this model, signal data without prior information is processed. This adaptive semi-supervised deep clustering algorithm not only improves the data utilization rate but also enhances the generalization ability and classification performance of the model.

Claims

1. A method for gait recognition in public safety areas based on LoRa, characterized in that: The method is implemented by the following steps: Step 1: Gait signal collection and preprocessing: Delete the null values ​​in the original LoRa signal, and convert the complex LoRa signal into a time series signal in the real domain. Then use the differential signal processing method to process the signal to highlight the signal changes caused by gait. Step 2: Gait feature extraction: reshape and convert the differential signal into image format to meet the model training requirements, input the converted image and corresponding labels into the convolutional neural network for training, and realize the effective classification of gait signals; Step 3: Adaptive semi-supervised deep clustering algorithm: The convolutional neural network is trained using an adaptive semi-supervised deep clustering algorithm, and the trained convolutional neural network is used for gait recognition.

2. A method for gait recognition in a public safety area based on LoRa according to claim 1, characterized in that: The specific process of step 1 is as follows: Step 1.1: Collect the LoRa signals of the two antennas at the same time. The LoRa signal collected by one antenna is an active LoRa signal, and the LoRa signal collected by the other antenna is an inactive LoRa signal. The LoRa signal is represented as a complex number, and the collected LoRa signal is read; Step 1.2: Delete the null value in the LoRa signal after reading; Step 1.3: Convert the complex number into the real number domain by extracting the real part, thereby generating a continuous time series signal sequence; Step 1.4: Using differential signal processing, the timing signal sequence formed by the active LoRa signal is subtracted from the timing signal sequence formed by the inactive LoRa signal to form a differential signal.

3. A method for gait recognition in a public safety area based on LoRa according to claim 2, characterized in that: The specific process of step 2 is as follows: Step 2.1: Extract key features: Reshape the differential signal, flatten the differential signal into a one-dimensional linear array, and fill or truncate it, normalize the reshaped data, scale it to a standard numerical range, and convert the normalized data into an image format; Step 2.2: Assign a corresponding label to each image data sample: Let the image data set be data, and its corresponding label set be labels. If labels[i] is equal to 1, it means that the label of the i-th image data sample is 0. Through this setting, the data set is divided into a labeled data set truedata and an unlabeled data set Nolabeldata. After the preparation is completed, the two data sets are input into the convolutional neural network for training so that the model can learn and identify the characteristics of different categories of data, thereby achieving effective classification.

4. The method for gait recognition in a public safety area based on LoRa according to claim 3, characterized in that: The specific process of step 3 is as follows: Step 3.1: Input the labeled and unlabeled datasets into the convolutional neural network for training; Step 3.2: Select data points in the unlabeled data set as the initial cluster centers. For each data point in the data set, calculate its distance to each cluster center and assign it to the cluster center with the closest distance. Step 3.3: For each cluster, recalculate its cluster center. The new cluster center is the mean of all data points in the cluster. The calculation formula is as follows: Where: S i is the data set of the i-th cluster center, |S i | is the number of data points in the set. This step is to repeat the allocation and update steps until the termination condition is met, that is, the distance between the new cluster center and the old cluster center is less than the preset threshold. At this time, each unlabeled data has a possible confidence category around a cluster center. By calculating the similarity between each unlabeled data and the corresponding cluster center and normalizing it to the range of [0, 1], data with different confidence levels are obtained. Step 3.4: In each iteration, an optimal threshold is selected, which divides the non-confidence values ​​of the data into two groups with the largest inter-class variance. After traversing all possible thresholds, a segmentation threshold that can maximize the inter-class variance is determined; then, for each category, if the number of samples with confidence lower than the threshold exceeds 10, the updated threshold is searched for in the same way. Finally, pseudo-label samples with high confidence are selected for each category, and these samples are saved. A dataset with pseudo-labels is created by setting a confidence threshold, and it is put into the model to expand the dataset; specifically, unlabeled data with confidence higher than the set threshold is regarded as pseudo-label data, and this data is input into CNN together with labeled data for the next round of training iteration. Through this process, the classification accuracy of the model is continuously improved.

5. The method for gait recognition in a public safety area based on LoRa according to claim 3, characterized in that: In step 3.2, Euclidean distance is used as the distance metric, and the calculation formula is as follows: Where x is the data, c i is the i-th cluster center, d is the data dimension, x j , c ij are x and c respectively i The value in the jth dimension.

Images