One-dimensional data target detection method and device

Abstract

The application relates to a label prediction network model training method and a one-dimensional data target detection method. The label prediction network model training method comprises the following steps: determining a plurality of target frames of one-dimensional training data, wherein the target frame is a meaningful data segment in the one-dimensional training data; calculating a three-line label formed by a left offset, a right offset and a foreground probability for each data point in each target frame; inputting the one-dimensional training data and the three-line label corresponding to each data point of each target frame into a label prediction network model for training, and obtaining a trained label prediction network model. The label prediction network model training method has the advantages that the three-line label coding and decoding technology is used to train the neural network model, the data length does not need to be concerned, various types of spectrum graph information can be comprehensively used for automatic feature extraction, and the method has strong real-time performance and high speed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, in particular to a one-dimensional data target detection method and device. BACKGROUND

[0002] One-dimensional data is a relatively common data form, such as WiFi-based CSI (Channel State Information) signal, voice signal, and electrocardiogram signal, and how to efficiently process one-dimensional data is an important research in machine learning. Generally, the processing of one-dimensional data is usually to obtain meaningful segments in the data and then perform subsequent processing. Taking activity recognition based on WiFi-based CSI signal as an example, the segments containing activities in the signal are first detected and extracted, and then recognition processing is performed. Therefore, the accuracy of detecting the segments containing activities in the signal affects the performance of subsequent recognition and classification tasks.

[0003] Most of the existing one-dimensional data target detection methods are based on threshold detection, in addition to which there are also detection methods based on deep learning. The principle of the threshold-based detection method is that the signal fluctuation range of the data containing activity segments is much larger than the fluctuation range of the data without activity, so a threshold is set in advance. If the range of data fluctuation exceeds the threshold, it means that activity has occurred. In this method, the size of the threshold directly affects the division result. However, the threshold in this method is often determined by subjective observation or past experience. As a result, the system may have performance degradation problems due to the involvement of human experience. In the detection method based on deep learning, there is a method that uses a state inference model to realize activity detection of one-dimensional data. The specific steps are first to discretize the one-dimensional data into containers of the same size, i.e., to periodically segment the one-dimensional data, then to use the state inference model to infer the state of each period, and finally to determine the starting point and ending point of the detected activity according to the state label, and to optimize the model feedback using the results of subsequent activity classification, thereby improving the accuracy of state inference. Since the system contains multiple network architectures, the system has strong robustness, but also because of this, it has the problem of complex structure, which leads to weak real-time performance and poor portability of the system. In addition, the size of the one-dimensional data segmentation period will affect whether the division data boundary is accurate. If the period is too large, the entire activity will fall into the period and state inference cannot be performed. If the period is too small, the system will judge some noise data as having activity significance. Therefore, the selection of the period size also affects the performance of the detection method. SUMMARY

[0004] In order to overcome at least one deficiency in the prior art, the embodiments of the present application provide a one-dimensional data target detection method and device.

[0005] In a first aspect, a training method of a label prediction network model is provided, comprising:

[0006] determining a plurality of target boxes of the one-dimensional training data, the target box being a meaningful data segment in the one-dimensional training data;

[0007] for each data point in each target box, calculating a distance between the data point and a start point of the target box, denoted as a left offset, calculating a distance between the data point and an end point of the target box, denoted as a right offset, and calculating a probability that the data point is located at a middle position of the target box, denoted as a foreground probability, the left offset, the right offset and the foreground probability forming a three-line label;

[0008] inputting the one-dimensional training data and the three-line label corresponding to each data point of each target box into the label prediction network model for training, to obtain a trained label prediction network model.

[0009] In an embodiment, the label prediction network model comprises a scaling convolutional layer, a feature transformation block and an inverse scaling convolutional layer.

[0010] The scaling convolutional layer is configured to perform feature extraction on the three-line label corresponding to each data point of each target box of the one-dimensional training data, to obtain a feature extraction result in the form of a two-dimensional tensor.

[0011] The feature transformation block is configured to process the feature extraction result in the form of a two-dimensional tensor, to obtain a class spectrum corresponding to the left offset, the right offset and the foreground probability respectively.

[0012] The inverse scaling convolutional layer is configured to process the class spectrum corresponding to the left offset, the right offset and the foreground probability respectively, to obtain a label prediction result corresponding to each data point of each target box, the label prediction result comprising predicted values of the left offset, the right offset and the foreground probability.

[0013] In an embodiment, the method further comprises:

[0014] decoding the label prediction result corresponding to each data point of each target box, to determine a target detection result of each target box, the target detection result comprising predicted values of the start point and the end point of the target box;

[0015] evaluating the label prediction network model obtained by training according to the predicted values of the start point and the end point of the target box and actual values of the start point and the end point of the target box, to obtain an evaluation result; the evaluation result is used to adjust parameters of the label prediction network model to train the label prediction network model.

[0016] In an embodiment, decoding the label prediction result corresponding to each data point of each target box, to determine a target detection result of each target box, comprises:

[0017] record the target frame currently being processed as a current target frame, and determine a maximum value of a predicted value of a foreground probability in a label prediction result corresponding to a data point in the current target frame;

[0018] determine whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value;

[0019] If the maximum value of the predicted value of the foreground probability is greater than the set threshold value, determine a target detection result of the current target frame according to a position of the data point corresponding to the maximum value of the predicted value of the foreground probability, a predicted value of a left offset and a predicted value of a right offset.

[0020] If the maximum value of the foreground probability is less than the set threshold value, process a next target frame.

[0021] In a second aspect, a one-dimensional data target detection method is provided, including:

[0022] input the one-dimensional data to be detected into a label prediction network model, and output a label prediction result corresponding to each data point of each target frame;

[0023] decode the label prediction result corresponding to each data point of each target frame, and determine a target detection result of each target frame, the target detection result including a predicted value of a start point and a predicted value of an end point of the target frame;

[0024] The label prediction network model is obtained according to the training method of the label prediction network model.

[0025] In an embodiment, decoding the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame includes:

[0026] record the target frame currently being processed as a current target frame, and determine a maximum value of a predicted value of a foreground probability in a label prediction result corresponding to a data point in the current target frame;

[0027] determine whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value;

[0028] If the maximum value of the predicted value of the foreground probability is greater than the set threshold value, determine a target detection result of the current target frame according to a position of the data point corresponding to the maximum value of the predicted value of the foreground probability, a predicted value of a left offset and a predicted value of a right offset.

[0029] If the maximum value of the foreground probability is less than the set threshold value, process a next target frame.

[0030] In a third aspect, a training device of a label prediction network model is provided, including:

[0031] The target bounding box determination module is used to determine multiple target bounding boxes in one-dimensional training data. The target bounding box is a meaningful data segment in the one-dimensional training data.

[0032] The three-line label calculation module is used to calculate the distance between the data point and the starting point of the target box for each data point in each target box, denoted as the left offset; calculate the distance between the data point and the ending point of the target box, denoted as the right offset; and calculate the probability that the data point is located in the middle of the target box, denoted as the foreground probability. The left offset, right offset, and foreground probability form the three-line label.

[0033] The model training module is used to input one-dimensional training data and the three-line label corresponding to each data point of each target box into the label prediction network model for training, so as to obtain the trained label prediction network model.

[0034] In one embodiment, it also includes:

[0035] The decoding module is used to decode the label prediction results corresponding to each data point of each target box, and determine the target detection result of each target box. The target detection result includes the predicted values ​​of the start and end points of the target box.

[0036] The evaluation module is used to evaluate the trained label prediction network model based on the predicted values ​​of the start and end points of the target boxes and the actual values ​​of the start and end points of the target boxes, and obtain the evaluation results. The evaluation results are used to adjust the parameters of the label prediction network model to train the label prediction network model.

[0037] Fourthly, a one-dimensional data target detection device is provided, comprising:

[0038] The label prediction module is used to input the one-dimensional data to be detected into the label prediction network model and output the label prediction result corresponding to each data point of each target box.

[0039] The decoding module is used to decode the label prediction results corresponding to each data point of each target box, and determine the target detection result of each target box. The target detection result includes the predicted values ​​of the start and end points of the target box.

[0040] The label prediction network model is obtained from the training device of the aforementioned label prediction network model.

[0041] In one embodiment, the decoding module is further configured to:

[0042] The target box currently being processed is recorded as the current target box, and the maximum predicted value of the foreground probability in the label prediction results corresponding to the data points in the current target box is determined;

[0043] Determine whether the maximum value of the predicted prospect probability is greater than a set threshold;

[0044] If the maximum of the prediction value of the foreground probability is greater than the set threshold value, a target detection result of the current target frame is determined according to the position of the data point corresponding to the maximum of the prediction value of the foreground probability, the prediction value of the left offset and the prediction value of the right offset.

[0045] If the maximum of the foreground probability is less than the set threshold value, the next target frame is processed.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] (1) The present application can complete pure data-driven feature analysis by processing data through a neural network, overcoming the problem of possible performance decline caused by deviation of the threshold value set by artificial experience in the threshold-based detection method;

[0048] (2) The present application can process signal segments of different lengths without paying attention to the specific length of the data by using a pure convolution structure of a scaling convolution layer, a full convolution network and an inverse scaling convolution layer, overcoming the problem of performance decline caused by periodic division of the signal in the existing detection method;

[0049] (3) The present application belongs to a single-stage anchor-free target detection method, does not need complicated feature engineering and is an end-to-end method, achieving good real-time performance and good migratability;

[0050] (4) The present application can effectively improve the accuracy and recall rate of detecting the interested segment in the signal, effectively improving the processing performance of the subsequent task and better providing the classifier of the subsequent recognition task with the interested segment containing the activity in the activity recognition based on the WIFI CSI signal.

[0051] In summary, the one-dimensional data target detection method of the present application does not need human experience, can cope with signal length changes, has a simple structure, strong real-time performance and good migratability. BRIEF DESCRIPTION OF DRAWINGS

[0052] The present application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which are included in the specification and form a part of the specification. In the drawings:

[0053] Figure 1 A flowchart of a training method of a label prediction network model according to an embodiment of the present application is shown;

[0054] Figure 2 A flowchart of a one-dimensional data target detection method according to an embodiment of the present application is shown;

[0055] Figure 3 A structural block diagram of a training device of a label prediction network model according to an embodiment of the present application is shown.

[0056] Figure 4 A structural block diagram of a one-dimensional data target detection device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0057] In the following, exemplary embodiments of the present application will be described with reference to the accompanying drawings. In the description, not all features of a practical embodiment are described for the sake of clarity and conciseness. It should be appreciated, however, that many embodiment-specific decisions can be made in the course of developing any such practical embodiment to achieve the specific objectives of the developer, and these decisions can vary from embodiment to embodiment.

[0058] It should also be noted here that, in order not to obscure the present application due to unnecessary details, only the device structures closely related to the scheme according to the present application are shown in the accompanying drawings, and other details not closely related to the present application are omitted.

[0059] It should be understood that the present application is not limited to the described embodiments by virtue of the following description with reference to the drawings. In this context, the embodiments can be combined with each other, features can be replaced or borrowed between different embodiments, and one or more features can be omitted in an embodiment.

[0060] Embodiments of the present application aim at the problems of manual participation, period segmentation, poor real-time performance, and difficulty in migration in the existing one-dimensional data target detection process. A neural network model is trained using three-line label encoding and decoding technology, and one-dimensional data to be detected is input into the trained neural network model to output a target detection result. The target detection method of the present application does not need to pay attention to the data length and can automatically extract features by comprehensively using multiple types of spectrum information, and has strong real-time performance and fast speed.

[0061] Figure 1 A flow block diagram of a training method of a label prediction network model according to an embodiment of the present application is shown. The method comprises: step S110, determining a plurality of target boxes of one-dimensional training data, the target box being a meaningful data segment in the one-dimensional training data;

[0062] In this step, the one-dimensional training data can be CSI data. First, the one-dimensional training data can be preprocessed. The specific preprocessing process includes denoising and normalization of the one-dimensional training data. For example, a Butterworth filter can be used to remove background environmental noise, and a z-score method can be used to eliminate baseline and singular sample data in the data. In this step, the target box is a meaningful data segment, i.e., an active data segment, in the one-dimensional training data. The meaningful data segment includes a start point and an end point. The data points in the target box can be labeled.

[0063] Then, in step S120, for each data point in each target box, the distance between the data point and the start point of the target box is calculated, denoted as left offset, the distance between the data point and the end point of the target box is calculated, denoted as right offset, and the probability that the data point is located at the middle position of the target box is calculated, denoted as foreground probability. The left offset, the right offset, and the foreground probability form a three-line label.

[0064] In this step, the start point position of the target box is denoted as s, the end point position is denoted as e, the distance between the data point position i and the start point position s is denoted as left offset L, L = i-s; the distance between the data point position i and the end point position e is denoted as right offset R, R = e-i; and the probability that the data point is located at the middle position of the target box is denoted as foreground probability, which can be obtained by using a Gaussian probability calculation formula or other existing probability calculation formula, which is not specifically limited here.

[0065] Then, in step S130, the one-dimensional training data and the three-line label corresponding to each data point of each target box are input into the label prediction network model for training to obtain a trained label prediction network model.

[0066] In this embodiment, the three-line label coding and decoding technology is used to train the neural network model, which does not need to pay attention to the data length, can automatically extract features by comprehensively considering various class spectrum graph information, has strong real-time performance, and is fast.

[0067] In one embodiment, the label prediction network model includes a scaling convolution layer, a feature transformation block, and an inverse scaling convolution layer.

[0068] The scaling convolution layer is used to extract features of the one-dimensional training data and the three-line label corresponding to each data point of each target box to obtain a feature extraction result in the form of a two-dimensional tensor.

[0069] The feature transformation block is used to process the feature extraction result in the form of a two-dimensional tensor to obtain class spectrum graphs corresponding to the left offset, the right offset, and the foreground probability, respectively.

[0070] The de-scaling convolutional layer is used to process the class spectrum corresponding to the left offset, the right offset and the foreground probability respectively, to obtain the label prediction result corresponding to each data point of each target frame, and the label prediction result includes the predicted value of the left offset, the right offset and the foreground probability.

[0071] In this embodiment, the feature transformation block includes a full convolutional layer, a batch normalization layer and a ReLU activation function, and can output three class frequency feature maps corresponding to the left offset, the right offset and the foreground probability respectively; the de-scaling convolutional layer includes three sub-de-scaling convolutional layers, and each sub-de-scaling convolutional layer is used to process one of the three class frequency feature maps. Here, the specific processing process of the label prediction network model can refer to the related content in the patent publication document with the application number CN202110417190.7.

[0072] In this embodiment, by using the scaling convolutional layer, the full convolutional network and the pure convolutional structure of the de-scaling convolutional layer, different length signal segments can be processed, and the specific length of the data does not need to be concerned, and the problem that the performance is reduced due to the period division of the signal in the existing detection method is overcome.

[0073] In one embodiment, the training method of the label prediction network model further includes evaluating the trained label prediction network model, specifically including:

[0074] The label prediction result corresponding to each data point of each target frame is decoded to determine the target detection result of each target frame, and the target detection result includes the predicted value of the start point and the end point of the target frame.

[0075] According to the predicted value of the start point and the end point of the target frame and the actual value of the start point and the end point of the target frame, the trained label prediction network model is evaluated to obtain an evaluation result; the evaluation result is used to adjust the parameters of the label prediction network model to train the label prediction network model.

[0076] In one embodiment, the label prediction result corresponding to each data point of each target frame is decoded to determine the target detection result of each target frame, including:

[0077] The target frame currently being processed is recorded as a current target frame, and the maximum value of the predicted value of the foreground probability in the label prediction result corresponding to the data point in the current target frame is determined.

[0078] determining whether the maximum of the predicted value of the foreground probability is greater than a set threshold value; if the maximum of the predicted value of the foreground probability is greater than the set threshold value, determining the target detection result of the current target frame according to the position of the data point corresponding to the maximum of the predicted value of the foreground probability, the predicted value of the left offset and the predicted value of the right offset; if the maximum of the predicted value of the foreground probability is less than the set threshold value, processing the next target frame. The set threshold value is selected according to actual conditions, which is not specifically limited here.

[0079] In this embodiment, the predicted value of the starting point of the current target frame is s' = i - L', where i is the position of the data point corresponding to the maximum of the predicted value of the foreground probability, and L' is the predicted value of the left offset of the data point corresponding to the maximum of the predicted value of the foreground probability. The predicted value of the end point of the current target frame is e' = i + R', where R' is the predicted value of the right offset of the data point corresponding to the maximum of the predicted value of the foreground probability.

[0080] In other embodiments, the label prediction result corresponding to each data point of each target frame is decoded to obtain multiple target detection results of the target frame. In order to avoid the above situation, non-maximum suppression processing can be performed on the multiple target detection results to obtain the final target detection result.

[0081] Figure 2 A flowchart of a one-dimensional data target detection method according to an embodiment of the present application is shown, the method comprising: step S210, inputting the one-dimensional data to be detected into a label prediction network model to output the label prediction result corresponding to each data point of each target frame; in this step, the label prediction network model is obtained according to the training method of the label prediction network model described above; in this step, the one-dimensional training data can be preprocessed first, and the specific preprocessing process includes denoising and normalization of the one-dimensional training data. For example, a Butterworth filter can be used to remove background environmental noise, and a z-score method can be used to eliminate the baseline and singular sample data in the data.

[0082] Then, in step S220, the label prediction result corresponding to each data point of each target frame is decoded to determine the target detection result of each target frame, and the target detection result includes the predicted values of the starting point and the end point of the target frame.

[0083] In this embodiment, the data is processed by a neural network, which can complete pure data-driven feature analysis and overcome the possible performance degradation problem caused by the deviation of the threshold value set by artificial experience due to external changes in the threshold-based detection method. It belongs to a single-stage anchor-free target detection method and does not need complicated feature engineering, which is an end-to-end method and achieves good real-time performance and portability.

[0084] In one embodiment, the label prediction result corresponding to each data point of each target frame is decoded to determine a target detection result of each target frame, including:

[0085] A target frame currently being processed is denoted as a current target frame, and a maximum value of a predicted value of a foreground probability in the label prediction result corresponding to the data point in the current target frame is determined;

[0086] It is judged whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value; if the maximum value of the predicted value of the foreground probability is greater than the set threshold value, a target detection result of the current target frame is determined according to a position of the data point corresponding to the maximum value of the predicted value of the foreground probability, a predicted value of a left offset and a predicted value of a right offset; if the maximum value of the predicted value of the foreground probability is less than the set threshold value, a next target frame is processed.

[0087] In this embodiment, a predicted value of a start point of the current target frame is s′=i-L′, where i is the position of the data point corresponding to the maximum value of the predicted value of the foreground probability, and L′ is the predicted value of the left offset of the data point corresponding to the maximum value of the predicted value of the foreground probability; a predicted value of an end point of the current target frame is e′=i+R′, where R′ is the predicted value of the right offset of the data point corresponding to the maximum value of the predicted value of the foreground probability.

[0088] Based on the same inventive concept as the training method of the label prediction network model provided in the embodiments of the present application, the present application also provides a training device of a label prediction network model corresponding thereto, Figure 3 A structural block diagram of the training device of the label prediction network model according to the embodiments of the present application is shown, and the device includes:

[0089] A target frame determination module 310 is configured to determine a plurality of target frames of one-dimensional training data, and the target frame is a meaningful data segment in the one-dimensional training data;

[0090] A three-line label calculation module 320 is configured to, for each data point in each target frame, calculate a distance between the data point and a start point of the target frame, denoted as a left offset, calculate a distance between the data point and an end point of the target frame, denoted as a right offset, and calculate a probability that the data point is located at a middle position of the target frame, denoted as a foreground probability, so as to form a three-line label of the left offset, the right offset and the foreground probability.

[0091] A model training module 330 is configured to input the one-dimensional training data and the three-line label corresponding to each data point of each target frame into the label prediction network model for training to obtain a trained label prediction network model.

[0092] In this embodiment, the three-line label encoding and decoding technology is used to train the neural network model, without the need to focus on the data length, and the automatic feature extraction can be performed on the comprehensive multi-class spectrum graph information, so that the real-time performance is strong and the speed is fast.

[0093] In an embodiment, the apparatus further comprises:

[0094] a decoding module configured to decode the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame, the target detection result comprising a predicted value of a start point and an end point of the target frame;

[0095] an evaluation module configured to evaluate the trained label prediction network model according to the predicted value of the start point and the end point of the target frame and an actual value of the start point and the end point of the target frame to obtain an evaluation result, the evaluation result being used to adjust parameters of the label prediction network model to train the label prediction network model.

[0096] Based on the same inventive concept as the one-dimensional data target detection method provided in the embodiments of the present application, the present application also provides a one-dimensional data target detection apparatus corresponding thereto, Figure 4 a structure block diagram of a one-dimensional data target detection apparatus according to an embodiment of the present application is shown, the apparatus comprising:

[0097] a label result prediction module 410 configured to input the one-dimensional data to be detected into a label prediction network model to output a label prediction result corresponding to each data point of each target frame, the label prediction network model being obtained according to the training apparatus of the label prediction network model described above;

[0098] a decoding module 420 configured to decode the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame, the target detection result comprising a predicted value of a start point and an end point of the target frame.

[0099] In this embodiment, the data is processed by the neural network, which can complete the feature analysis driven by pure data, and overcome the possible performance decline problem caused by the deviation of the threshold value set by artificial experience due to external changes in the threshold-based detection method; it is a single-stage anchor-free target detection method, does not need to perform complex feature engineering, and is an end-to-end method, which realizes good real-time performance and migratability.

[0100] In an embodiment, the decoding module 420 is further configured to:

[0101] record the target frame currently being processed as a current target frame, and determine a maximum value of the predicted value of the foreground probability in the label prediction result corresponding to the data point in the current target frame;

[0102] determine whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value;

[0103] If the maximum of the prediction value of the foreground probability is greater than a set threshold, a target detection result of the current target frame is determined according to a position of a data point corresponding to the maximum of the prediction value of the foreground probability, the prediction value of the left offset and the prediction value of the right offset.

[0104] If the maximum of the foreground probability is less than the set threshold, a next target frame is processed.

[0105] In summary, in the one-dimensional data target detection process, the method of the present application does not need human experience, can cope with signal length changes, has simple structure, strong real-time performance and good migration performance.

[0106] In order to demonstrate the feasibility of the one-dimensional data target detection method of the present application, experimental analysis is carried out by using data collected independently. The independently collected data contains 1200 detected different behavior segments of CSI data, including gesture, walking and running data, etc., a total of 12 different behaviors. When the model is trained, 80% of the data is selected as the training set and 20% as the test set.

[0107] Through experimental analysis, the method proposed in the present application can correctly detect events in a number of long sequence signal data and obtain the start and end positions of multiple events; by manual marking, different types of signal data are adapted to achieve good detection effect. In the data collected in this experiment, the recall rate of the method proposed in the present application on the long sequence CSI signal of different behavior activities is 99.6%, which is 3.4% higher than that of the previous method; the IoU (Intersection over Union) is 86.2%, which is 2.3% higher than that of the previous method, wherein the recall rate is the proportion of the true target that is predicted out, and the IoU is the intersection over union of the predicted range and the actual range.

[0108] In the embodiments of the present disclosure, it should be understood that the disclosed apparatus and method can also be implemented in other manners. The embodiments described above are merely exemplary. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architectures, functions and operation of the apparatus, method and computer program product according to the embodiments of the present disclosure. In this regard, each block in the flowcharts and block diagrams can represent a module, a segment or a portion of code which comprises one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the accompanying drawings. For example, two blocks noted in succession can in fact be executed substantially concurrently or in the reverse order, depending on the functionality involved. It should also be noted that each block in the block diagrams and / or flowcharts and combinations of blocks in the block diagrams and / or flowcharts can be implemented by dedicated hardware-based systems which perform the specified functions or actions, or can be implemented by a combination of dedicated hardware-based systems and computer instructions.

[0109] In addition, the various functional modules in the embodiments of the present disclosure can be integrated together to form a separate part, or can exist independently, or two or more modules can be integrated to form a separate part.

[0110] If the functions are implemented in the form of software function modules and sold or used as an independent product, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present disclosure can essentially or contribute to the prior art, or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present disclosure. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0111] The above describes only various embodiments of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present disclosure, which should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A method for training a label prediction network model, characterized in that, The method comprises the following steps: determining a plurality of target frames of one-dimensional training data, the target frames being meaningful data segments in the one-dimensional training data; the one-dimensional training data is a WIFI CSI signal; for each data point in each of the target frames, calculating the distance between the data point and the starting point of the target frame, denoted as left offset, calculating the distance between the data point and the ending point of the target frame, denoted as right offset, and calculating the probability that the data point is located at the middle position of the target frame, denoted as foreground probability, the left offset, the right offset and the foreground probability form a three-line label; inputting the one-dimensional training data and the three-line label corresponding to each data point of each target frame into a label prediction network model for training to obtain a trained label prediction network model; the label prediction network model comprises a scaling convolution layer, a feature transformation block and an inverse scaling convolution layer; the scaling convolution layer is used for feature extraction of the one-dimensional training data and the three-line label corresponding to each data point of each target frame to obtain a feature extraction result in the form of a two-dimensional tensor; the feature transformation block is used for processing the feature extraction result in the form of a two-dimensional tensor to obtain a class spectrum diagram corresponding to the left offset, the right offset and the foreground probability respectively; the inverse scaling convolution layer is used for processing the class spectrum diagram corresponding to the left offset, the right offset and the foreground probability respectively to obtain a label prediction result corresponding to each data point of each target frame, the label prediction result comprising predicted values of the left offset, the right offset and the foreground probability.

2. The method of claim 1, wherein, The method further comprises: decoding the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame, the target detection result comprising predicted values of the starting point and the ending point of the target frame; evaluating the trained label prediction network model according to the predicted values of the starting point and the ending point of the target frame and the actual values of the starting point and the ending point of the target frame to obtain an evaluation result; the evaluation result is used to adjust the parameters of the label prediction network model to train the label prediction network model.

3. The method of claim 2, wherein, wherein, decoding the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame comprises: denoting the target frame currently being processed as a current target frame, and determining the maximum value of the predicted value of the foreground probability in the label prediction result corresponding to the data point in the current target frame; determining whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value; if the maximum value of the predicted value of the foreground probability is greater than the set threshold value, determining the target detection result of the current target frame according to the position of the data point corresponding to the maximum value of the predicted value of the foreground probability, the predicted value of the left offset and the predicted value of the right offset; if the maximum value of the foreground probability is less than the set threshold value, processing the next target frame.

4. A method of detecting a one-dimensional data target, characterized by The method comprises the following steps: inputting the one-dimensional data to be detected into the label prediction network model to output a label prediction result corresponding to each data point of each target frame; decoding the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame, the target detection result including predicted values of the start point and the end point of the target frame; The label prediction network model is obtained by the training method of the label prediction network model according to any one of claims 1-3.

5. The method of claim 4, wherein, Wherein, decoding the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame, including: record the target frame currently being processed as a current target frame, and determine the maximum value of the predicted value of the foreground probability in the label prediction result corresponding to the data point in the current target frame; determine whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value; if the maximum value of the predicted value of the foreground probability is greater than the set threshold value, determine the target detection result of the current target frame according to the position of the data point corresponding to the maximum value of the predicted value of the foreground probability, the predicted value of the left offset and the predicted value of the right offset; if the maximum value of the foreground probability is less than the set threshold value, process the next target frame. 6.A device for training a label prediction network model, characterized in that, including: a target frame determination module configured to determine a plurality of target frames of one-dimensional training data, the target frame being a meaningful data segment in the one-dimensional training data; the one-dimensional training data is a WIFI CSI signal; a three-line label calculation module configured to, for each data point in each target frame, calculate a distance between the data point and a start point of the target frame, denoted as a left offset, calculate a distance between the data point and an end point of the target frame, denoted as a right offset, and calculate a probability that the data point is located at a middle position of the target frame, denoted as a foreground probability, the left offset, the right offset and the foreground probability forming a three-line label; a model training module configured to input the one-dimensional training data and the three-line label corresponding to each data point of each target frame into a label prediction network model for training to obtain a trained label prediction network model; the label prediction network model includes a scaling convolution layer, a feature transformation block and an inverse scaling convolution layer; the scaling convolution layer is configured to perform feature extraction on the one-dimensional training data and the three-line label corresponding to each data point of each target frame to obtain a feature extraction result in the form of a two-dimensional tensor; the feature transformation block is configured to process the feature extraction result in the form of a two-dimensional tensor to obtain a class spectrum diagram corresponding to the left offset, the right offset and the foreground probability respectively; the inverse scaling convolution layer is configured to process the class spectrum diagram corresponding to the left offset, the right offset and the foreground probability respectively to obtain a label prediction result corresponding to each data point of each target frame, the label prediction result including the predicted values of the left offset, the right offset and the foreground probability.

7. The apparatus of claim 6, wherein, Further comprising: a decoding module configured to decode the label prediction result corresponding to each data point of each target frame to determine a target detection result of each target frame, the target detection result including predicted values of the start point and the end point of the target frame; An evaluation module is configured to evaluate the trained label prediction network model according to the predicted values of the start point and the end point of the target frame and the actual values of the start point and the end point of the target frame, and obtain an evaluation result; the evaluation result is used to adjust the parameters of the label prediction network model to train the label prediction network model.

8. A one-dimensional data target detection apparatus, characterized by comprising: Comprise: A label result prediction module is configured to input one-dimensional data to be detected into a label prediction network model, and output a label prediction result corresponding to each data point of each target frame; A decoding module is configured to decode the label prediction result corresponding to each data point of each target frame, and determine a target detection result of each target frame, wherein the target detection result comprises predicted values of the start point and the end point of the target frame. The label prediction network model is obtained by the training device of the label prediction network model according to claim 6.

9. The apparatus of claim 8, wherein, The decoding module is further configured to: Record the target frame currently being processed as a current target frame, and determine a maximum value of the predicted value of the foreground probability in the label prediction result corresponding to the data point in the current target frame; Judge whether the maximum value of the predicted value of the foreground probability is greater than a set threshold value; If the maximum value of the predicted value of the foreground probability is greater than the set threshold value, determine the target detection result of the current target frame according to the position of the data point corresponding to the maximum value of the predicted value of the foreground probability, the predicted value of the left offset and the predicted value of the right offset; If the maximum value of the foreground probability is less than the set threshold value, process the next target frame.

Images