Class identification method, electronic device and program product

Abstract

The embodiment of the invention provides a category identification method, electronic equipment and a program product. The category identification method comprises the following steps: acquiring a target point cloud corresponding to a to-be-identified target; extracting current target features based on the physical measurement quantity of the target point cloud; determining current category probability distribution based on the current target features and historical target features of the to-be-recognized target; and determining the category of the to-be-identified target based on the current category probability distribution and the historical category probability distribution of the to-be-identified target. Through the scheme of the embodiment, the calculation overhead can be reduced, the data processing efficiency can be improved, the behavior trend of the target is comprehensively considered, the influence caused by inaccurate single measurement is reduced, and the decision accuracy and robustness of the system are improved.

Description

Technical Field

[0001] This disclosure relates to the field of integrated communication and sensing technology, and in particular to a category identification method, electronic device, and program product. Background Technology

[0002] In recent years, with the continuous development of radar technology, integrating target recognition into the traditional target detection process to improve the overall performance of radar systems has become one of the development trends in the field of radar target detection. However, radar target detection technology mainly focuses on detecting the presence of targets, while its ability to identify the type of target to be identified is relatively weak.

[0003] Related technologies for target feature extraction and recognition mainly include: micro-Doppler-based feature extraction and target recognition methods and single-snapshot point cloud recognition methods based on traditional machine learning. However, micro-Doppler-based feature extraction and target recognition methods suffer from problems such as high computational overhead, poor real-time performance, and susceptibility to external environmental influences leading to performance degradation. Traditional machine learning-based single-snapshot point cloud recognition methods involve arduous manual feature engineering requiring deep expertise, are time-consuming and labor-intensive, are prone to introducing human bias and errors during feature design, and struggle to extract high-order features and complex relationships from the data. Summary of the Invention

[0004] This disclosure provides a category identification method, electronic device, and program product.

[0005] In a first aspect, embodiments of this disclosure provide a category identification method, which includes:

[0006] Obtain the target point cloud corresponding to the target to be identified;

[0007] Extract the current target features based on the physical measurements of the target point cloud;

[0008] The current category probability distribution is determined based on the current target features and the historical target features of the target to be identified;

[0009] The category of the target to be identified is determined based on the current category probability distribution and the historical category probability distribution of the target to be identified.

[0010] Secondly, embodiments of this disclosure also provide an electronic device, including:

[0011] One or more processors;

[0012] A memory having stored one or more programs that, when executed by one or more processors, enable the one or more processors to implement the category recognition method.

[0013] One or more input / output (I / O) interfaces are connected between the processor and the memory and configured to enable information exchange between the processor and the memory.

[0014] Thirdly, this disclosure also provides a computer program product, which includes a computer program that, when executed by a processor, implements the category recognition method.

[0015] This embodiment of the scheme acquires a target point cloud corresponding to the target to be identified; extracts current target features based on physical measurements of the target point cloud; determines the current category probability distribution based on the current target features and historical target features of the target to be identified; and determines the category of the target to be identified based on the current category probability distribution and historical category probability distribution of the target to be identified. This embodiment reduces computational overhead, improves data processing efficiency, comprehensively considers the behavioral trends of the target, reduces the impact of inaccurate single measurements, and improves the decision-making accuracy and robustness of the system. Attached Figure Description

[0016] In the accompanying drawings of the embodiments disclosed herein:

[0017] Figure 1 This is a schematic flowchart of the category identification method provided in the embodiments of this disclosure;

[0018] Figure 2 A schematic diagram illustrating the current category probability distribution determination method based on an LSTM model provided in this embodiment of the disclosure;

[0019] Figure 3 A schematic diagram of the current category probability distribution determination method based on LSTM model provided in this embodiment of the disclosure;

[0020] Figure 4 A flowchart of the reasoning result acquisition method provided in this embodiment of the disclosure;

[0021] Figure 5 This is a flowchart of a method for obtaining latent features corresponding to current target features provided in an embodiment of this disclosure;

[0022] Figure 6 A flowchart illustrating the training method for a neural network model provided in this embodiment of the disclosure;

[0023] Figure 7 The flowchart of the method for obtaining the current sample features and historical sample features of the target sample provided in this embodiment of the disclosure constitutes a training sample;

[0024] Figure 8 A flowchart of a method for sampling the motion trajectory of a target sample based on a preset sliding window to obtain basic training samples, provided in an embodiment of this disclosure;

[0025] Figure 9 This is a schematic diagram of a method for sampling target motion sub-trajectories based on a preset sliding window, provided in an embodiment of this disclosure.

[0026] Figure 10 A flowchart of a method for training a preset neural network based on training samples to obtain a neural network model, provided in an embodiment of this disclosure;

[0027] Figure 11 A flowchart illustrating a method for obtaining the difference between the category of a sample target predicted by a neural network and the category labeled on the sample target, as provided in this embodiment of the disclosure;

[0028] Figure 12 A schematic flowchart illustrating a method for determining the category of a target to be identified based on the current category probability distribution and the historical category probability distribution of the target to be identified, as provided in this embodiment of the disclosure;

[0029] Figure 13 This is a schematic diagram illustrating the sum of the current category probability distribution and the historical category probability distribution provided in this embodiment of the disclosure;

[0030] Figure 14 A flowchart illustrating a method for determining the category of a target to be identified based on the confidence level of each category, as provided in embodiments of this disclosure.

[0031] Figure 15 This is a schematic diagram of the electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions of this disclosure, the communication-sensing data processing method and computer-readable storage medium provided in the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings.

[0033] The present disclosure will be described more fully below with reference to the accompanying drawings; however, the embodiments shown may be embodied in different forms, and the present disclosure should not be construed as limited to the embodiments set forth below. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will enable those skilled in the art to fully understand the scope of the disclosure.

[0034] The accompanying drawings of the embodiments disclosed herein are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the detailed embodiments to explain this disclosure and do not constitute a limitation thereof. The above and other features and advantages will become more apparent to those skilled in the art from the description of the detailed embodiments with reference to the accompanying drawings.

[0035] This disclosure may be described with reference to plan and / or cross-sectional views using the ideal schematic diagrams of this disclosure. Therefore, the example illustrations may be modified according to manufacturing techniques and / or tolerances.

[0036] Where there is no conflict, the various embodiments of this disclosure and the features thereof in the embodiments may be combined with each other.

[0037] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the disclosure. The term "and / or" as used in this disclosure includes any and all combinations of one or more of the associated enumerated entries. The singular forms "a" and "the" as used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. The terms "comprising," "made of," etc., as used in this disclosure specify the presence of the stated feature, integral, step, operation, element, and / or component, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof.

[0038] Unless otherwise specified, all terms used in this disclosure (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined in this disclosure.

[0039] In recent years, with the continuous development of radar technology, integrating target recognition into the traditional target detection process to improve the overall performance of radar systems has become one of the development trends in the field of radar target detection. However, radar target detection technology mainly focuses on detecting the presence of targets, while its ability to identify target types is relatively weak. Because radar echo signals contain various information from targets, clutter, and noise, accurately determining targets becomes difficult. Especially in situations with low signal-to-noise ratios, multiple targets, and complex environments, recognition performance often degrades significantly. A typical radar target recognition process includes steps such as data acquisition (collecting raw radar echo data, including target echoes, clutter, and noise), preprocessing (correcting and enhancing the raw data to remove noise and interference), feature extraction (constructing target features for decision-making), and final classification decision.

[0040] The target feature extraction and recognition techniques in related technologies mainly use the following schemes:

[0041] Option 1: Feature Extraction and Target Recognition Method Based on Micro-Doppler

[0042] The micro-Doppler effect in radar refers to the Doppler frequency shift in the radar echo caused by the radial motion or rotation of a target. This effect contains information about the target's motion and scattering characteristics, and therefore serves as an important basis for target identification in many studies. A commonly used micro-Doppler feature extraction method is based on time-frequency analysis. This method converts the radar echo signal into the time-frequency domain and then extracts micro-Doppler features by analyzing the time-frequency spectrum. A commonly used time-frequency analysis method is the Short-Time Fourier Transform (STFT). The STFT divides the signal into short-time windows and applies a Fourier transform to each window to obtain the time-frequency spectrum within that window. After obtaining the target's micro-Doppler features, a common identification method is to use a deep learning model based on a Convolutional Neural Network (CNN) for target classification.

[0043] Disadvantages of Option 1: The main disadvantages of this method are as follows:

[0044] High computational overhead: Micro-Doppler-based feature extraction and target recognition methods typically require a large amount of computation, especially when performing time-frequency analysis and sliding window STFT. This requires processing a large amount of data and parameters, and when there are many targets, it is necessary to calculate the micro-Doppler features of each target individually, which leads to extremely high computational overhead.

[0045] Poor real-time performance: Due to the high computational cost and complexity, these methods face challenges in real-time performance, especially in scenarios requiring real-time target recognition, where they cannot meet real-time requirements.

[0046] Susceptible to external environmental influences: Micro-Doppler feature extraction methods are sensitive to clutter and noise, making them susceptible to external environmental influences and leading to decreased recognition performance. For example, in low signal-to-noise ratio environments, micro-Doppler features can be overwhelmed by clutter and noise, making it difficult to extract effective target features.

[0047] Option 2: Single-snapshot point cloud recognition method based on traditional machine learning

[0048] In this method, a series of manual feature engineering operations are required for point cloud data captured in a single snapshot. For example, to extract features based on the geometric shape of the target, relevant parameters of the geometric shape formed by the point cloud are calculated, such as the area, perimeter, and volume of the target point cloud. These features can reflect the size and shape information of the target to a certain extent. Simultaneously, intensity features are considered, such as calculating the average intensity and intensity variance of the point cloud to obtain the target's reflectivity characteristics. Furthermore, point cloud distribution features are also an important part of the manual feature engineering, such as calculating point cloud density and dispersion to understand the spatial distribution of the target point cloud. After feature extraction, traditional machine learning algorithms such as support vector machines and decision trees are typically chosen to build a classification model.

[0049] Option 2 has several drawbacks:

[0050] On the one hand, manual feature engineering is an extremely demanding task requiring deep expertise. It is not only time-consuming and labor-intensive, but also prone to introducing human bias and errors during feature design. Different types of targets and diverse application scenarios often require different combinations of features, making this method lack versatility and flexibility. For example, when facing new targets or complex environmental changes, features may need to be redesigned.

[0051] On the other hand, it is difficult to extract high-order features and complex relationships from the data. Due to limitations in human cognition and design capabilities, the extracted features often only reflect the surface information of the target, failing to effectively capture deeper target movement patterns, behavioral trends, and complex interactions between targets. For example, in dynamic scenes, high-order dynamic information such as changes in the target's trajectory and speed trends cannot be accurately represented by simple artificial features. This leads to a bottleneck in the performance improvement of classification models built based on these artificial features in complex targets and changing environments. The recognition accuracy is difficult to further improve, failing to meet the demand for high-precision target recognition in practical applications. It also proves inadequate in the face of increasingly complex real-world scenarios, limiting its widespread application and performance improvement potential in the field of target recognition.

[0052] This embodiment of the scheme acquires a target point cloud corresponding to the target to be identified; extracts current target features based on physical measurements of the target point cloud; determines the current category probability distribution based on the current target features and historical target features of the target to be identified; and determines the category of the target to be identified based on the current category probability distribution and historical category probability distribution of the target to be identified. This embodiment reduces computational overhead, improves data processing efficiency, comprehensively considers the behavioral trends of the target, reduces the impact of inaccurate single measurements, and improves the decision-making accuracy and robustness of the system.

[0053] The solutions disclosed herein can be applied to any target monitoring and identification scenario, including but not limited to: integrated communication and sensing technology scenarios, such as integrated communication and sensing scenarios requiring accurate target identification and classification, especially suitable for dynamic target identification scenarios in integrated communication and sensing base stations. They can be mainly applied to scenarios requiring real-time, accurate identification and classification of dynamic targets, and have significant application value under conditions with strict requirements for real-time performance and computational efficiency. The following are some application scenario examples:

[0054] 1. In intelligent transportation systems, the solutions of this disclosure can provide data support for traffic flow monitoring by analyzing vehicle trajectories in real time. For example, accurately counting the number of vehicles passing through different road sections at different time periods provides a basis for traffic management departments to adjust traffic light durations and optimize road planning. Simultaneously, it can effectively prevent traffic accidents. By identifying vehicle trajectories, abnormal driving behaviors of vehicles, such as sudden lane changes and emergency braking, can be detected in a timely manner, and early warnings can be issued to remind surrounding vehicles to pay attention to safety.

[0055] 2. In the field of urban security monitoring, the solution of this disclosure embodiment can be applied to the monitoring system to identify and classify dynamic targets such as people and vehicles within the monitoring range. By analyzing the behavior of the targets, abnormal behaviors, such as the gathering of people or the abnormal parking of vehicles, can be detected in a timely manner, and alarms can be issued to the monitoring personnel in a timely manner.

[0056] 3. In the environment surrounding airports, bird activity can pose a serious threat to aircraft takeoff and landing. The solution disclosed in this embodiment can be applied to airport bird monitoring systems. By analyzing the flight paths and characteristics of birds, different bird species can be quickly and accurately identified. This helps airport management take targeted bird control measures, such as playing calls of specific bird predators or using optical equipment to disrupt bird flight paths, thereby reducing the risk of bird collisions with aircraft and ensuring airport flight safety.

[0057] 4. In the field of aquatic environment monitoring and management, accurate identification of vessels in the water is crucial. The solution of this disclosure can be applied to aquatic vessel monitoring system. By analyzing the navigation trajectory and characteristics of vessels, different types of vessels can be identified. It can also determine the vessel's navigation direction, speed, and intention, providing data support for waterway regulatory departments to better maintain waterway order, such as preventing illegal fishing, maritime smuggling, and piracy.

[0058] 5. Drone Monitoring Scenarios: With the rapid development of the low-altitude economy, the use of drones is becoming increasingly widespread, but this also brings problems such as unauthorized flights and intrusions into sensitive areas. The solution disclosed in this embodiment can be applied to a drone identification and detection system. Through trajectory analysis and feature recognition, it accurately determines the type, flight direction, and intent of the drone. Whether over urban areas or in specific controlled areas, it can promptly detect drones flying illegally and take corresponding measures, such as issuing warnings or guiding them to land, ensuring low-altitude flight safety and public safety. In the military field, for some key areas, such as military bases, test ranges, and border defense lines, the military needs to conduct strict drone monitoring of the surrounding airspace. The solution disclosed in this embodiment can play an important role in this, quickly distinguishing between civilian drones and drones that may pose a potential threat through real-time monitoring and feature recognition. For drones that enter sensitive airspace without authorization, the system can promptly issue an alarm and provide their precise location and flight trajectory information, so that the military can take corresponding countermeasures, such as interfering with their communication links, guiding them away from the target area, or directly capturing them, ensuring the security and confidentiality of military areas.

[0059] The embodiments of this disclosure will be described in detail below.

[0060] This disclosure provides a category identification method, such as... Figure 1 As shown, it includes steps S11-S14:

[0061] S11. Obtain the target point cloud corresponding to the target to be identified.

[0062] In this embodiment of the disclosure, the target to be identified may include, but is not limited to, any object that needs to be detected or identified, such as ships, flying objects, people, animals, etc.

[0063] In this embodiment of the disclosure, the sensing signal receiver can receive and process the sensing signal so that subsequent steps can better extract target information (such as target point cloud). Processing the sensing signal may include, but is not limited to: filtering out possible target sensing signals from the sensing signal, for example, removing some interference information and background noise to make the target sensing signal more prominent; or, for example, distinguishing between target sensing signals and non-target sensing signals based on thresholds, feature matching, or other methods.

[0064] In this embodiment of the disclosure, the target sensing signal can be subjected to spectrum analysis and processing using digital signal processing techniques (such as Fourier transform). During this process, multiple snapshots can be taken. In each snapshot, the sensing signal receiver can obtain a point cloud of potentially interesting targets based on information such as the RD (Range-Doppler Map) map and angle estimation.

[0065] S12. Extract the current target features based on the physical measurements of the target point cloud.

[0066] In this embodiment of the disclosure, each point in the target point cloud carries physical measurements, which contain rich information about the target to be identified.

[0067] In this embodiment of the disclosure, the physical measurements of the target point cloud may include, but are not limited to, the spatial coordinates and / or echo power of each point in the target point cloud.

[0068] In this embodiment of the disclosure, spatial coordinates can be represented as Spatial coordinates can accurately locate the position of the target to be identified in three-dimensional space. Echo power can be expressed as... The echo power reflects the intensity of the reflected signal from the target to be identified.

[0069] In this embodiment of the disclosure, the current target features may include, but are not limited to, at least one of the following: the number of points in the target point cloud, the point cloud diameter, and the standard deviation of the echo power.

[0070] In this embodiment of the disclosure, the extraction of current target features based on physical measurements of the target point cloud may include, but is not limited to, at least one of the following:

[0071] Count the number of points in the target point cloud;

[0072] The point cloud diameter of the target point cloud is determined based on the spatial coordinates of each point; and,

[0073] The standard deviation of the echo power is calculated based on the echo power.

[0074] In this embodiment of the disclosure, the number of points in the target point cloud can also be called the cluster point number, which can be represented by N (N is a positive integer). Counting the number of points in a cluster of target point clouds can be used as a reference indicator of the size of the target to be identified.

[0075] In this embodiment of the disclosure, the point cloud diameter can be represented by d (d is a positive number). The pairwise distance between each point in the target point cloud can be calculated, and the maximum value of the calculated distance is the point cloud diameter. The point cloud diameter reflects the approximate size range of the target to be identified in space.

[0076] In this embodiment of the disclosure, the standard deviation of the echo power can be used This means that it is possible to target all points in the point cloud. The standard deviation of the echo power is obtained by calculating the standard deviation of the value. The standard deviation of the echo power reflects the fluctuation of the echo power within the target point cloud.

[0077] In this embodiment of the disclosure, the target features corresponding to the target point cloud obtained at snapshot t can be concatenated to obtain the target feature vector at snapshot t. : .in, This represents the diameter of the point cloud at time t. This represents the number of points at time t. This represents the standard deviation of the echo power at time t.

[0078] In this embodiment of the disclosure, after obtaining the target point cloud corresponding to the target to be identified, the method may further include:

[0079] Obtain unordered point cloud data of the target point cloud;

[0080] Input the disordered point cloud data into a preset point cloud neural network model;

[0081] Inference is performed on disordered point cloud data based on a point cloud neural network model to extract the features of the current target.

[0082] In the embodiments of this disclosure, the above scheme describes a simple feature extraction scheme. Feature extraction can also be performed using a preset point cloud neural network model. This point cloud neural network model may include, but is not limited to, PointNet (point cloud network, a deep learning model for point sets used for 3D classification and segmentation). That is, unordered target point cloud data is input into PointNet, and PointNet extracts effective global and local current target features.

[0083] S13. Determine the probability distribution of the current category based on the current target features and the historical target features of the target to be identified.

[0084] In this embodiment of the disclosure, the current target feature is the feature of the current cluster target point cloud corresponding to the target to be identified, and the historical target feature is the feature of at least one cluster of historical target point clouds corresponding to the target to be identified; the historical target point cloud is the target point cloud collected before the current cluster target point cloud.

[0085] In this embodiment of the disclosure, the current target feature may include the features of the target point cloud obtained at snapshot t, and the historical target feature may include the features of the target point cloud obtained before snapshot t (e.g., snapshot t-1).

[0086] In this embodiment of the disclosure, determining the current category probability distribution based on the current target features and the historical target features of the target to be identified may include:

[0087] Based on a pre-defined neural network model, reasoning is performed on the current target features and historical target features to obtain the probability distribution of the current category; this neural network model is capable of processing time series data and capturing the time dependencies between time series data.

[0088] In this embodiment of the disclosure, the neural network model may include, but is not limited to, LSTM (Long Short-Term Memory) model, RNN (Recurrent Neural Network) model, and Transformer model.

[0089] In this embodiment of the disclosure, the historical target features may include, but are not limited to: historical latent features and / or historical memory features obtained by reasoning from the historical target point cloud by a neural network model.

[0090] In this embodiment of the disclosure, the historical latent features may include the latent features of the target point cloud at snapshot t-1, which can be represented as historical latent vectors. Historical memory features can include the memory features of the target point cloud at snapshot time t-1, which can be represented as a historical memory vector. .

[0091] In the embodiments disclosed herein, such as Figure 2 As shown, taking the LSTM model as an example, the embodiment of this disclosure can be illustrated by describing the historical latent vector output by the LSTM model at the previous time step. and historical memory vector The target feature vector output by the LSTM model at the current time t Normalized target feature vector obtained after normalization When input into an LSTM model, the LSTM model can effectively process sequential data and capture long-term dependencies within the data. However, when inferring from the target point cloud of the first snapshot using a neural network model, the historical latent vector from the previous time step t-1 is unavailable. and historical memory vector The historical latent vector at this time and historical memory vector It can be represented by a vector of all zeros.

[0092] In the embodiments disclosed herein, such as Figure 3 As shown, based on a preset neural network model, reasoning is performed on the current target features and historical target features to obtain the probability distribution of the current category, which may include steps S21-S24:

[0093] S21. Preprocess the current target features to obtain the normalized target feature vector.

[0094] In this embodiment of the disclosure, normalizing the current target features can be performed on the feature vector. Normalization is performed to obtain the normalized target feature vector. (i.e., the normalized current target features), where the target feature vector The first in ( (where the number of features is a positive integer) can be represented as , target feature vector The first in Features The expression for normalization is:

[0095] ;

[0096] in, For the first Features The normalized value, and Yes, it is a configurable hyperparameter. Indicates the first Features The minimum value, Indicates the first Features The maximum value.

[0097] In this embodiment of the disclosure, the normalized target feature vector It can be derived from the target feature vector Normalized values ​​of each feature Composition. The target eigenvector at the previous time t-1 can be represented as: The normalized target eigenvector of the previous time step t-1 can be expressed as: (i.e., the historical target characteristics of normalization).

[0098] S22. Input the normalized target feature vector into the LSTM model. The LSTM model then obtains the inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent features.

[0099] In this embodiment of the disclosure, the reasoning result may include, but is not limited to, a first reasoning result, a second reasoning result, and a third reasoning result.

[0100] In the embodiments disclosed herein, such as Figure 4 As shown, the normalized target feature vector is input into the LSTM model, and the LSTM model obtains the inference result based on the normalized target feature vector and the historical latent vectors corresponding to the historical latent features, including steps S31-S33:

[0101] S31. Input the normalized target feature vector into the input gate of the LSTM model. The input gate obtains the first inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent features.

[0102] In this embodiment of the disclosure, the normalized target feature vector can be... Input the LSTM model into its input gate and obtain the first inference result based on the following calculation: ;in, This is the first reasoning result. The preset input gate feature weight parameters, The input gate latent vector weight parameters are preset. For the preset input gate bias term, , , It can be acquired through learning.

[0103] S32. Input the normalized target feature vector into the forget gate of the LSTM model. The forget gate obtains the second inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent features.

[0104] In this embodiment of the disclosure, the normalized target feature vector can be... Input the forget gate of the LSTM model and obtain the second inference result based on the following calculation formula: ;in, This is the result of the second inference. The preset forget gate feature weight parameters, The preset weight parameters for the latent vector of the forget gate. For the preset forget gate bias term, , , It can be acquired through learning.

[0105] S33. Input the normalized target feature vector into the output gate of the LSTM model. The output gate obtains the third inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent features.

[0106] In this embodiment of the disclosure, the normalized target feature vector can be... Input the output gate of the LSTM model and obtain the third inference result based on the following calculation formula: ;in, This is the third reasoning result. The preset output gate feature weight parameters, The preset output gate latent vector weight parameters, For the preset output gate bias term, , , It can be acquired through learning. It is the sigmoid function, and the formula is as follows: y represents the dependent variable of the sigmoid function, and x represents the independent variable of the sigmoid function.

[0107] S23. The LSTM model obtains the latent features corresponding to the current target features based on the inference results, historical latent vectors, and historical memory vectors corresponding to historical memory features.

[0108] In the embodiments disclosed herein, such as Figure 5 As shown, the LSTM model obtains the latent features corresponding to the current target feature based on the inference results, historical latent vectors, and historical memory vectors corresponding to historical memory features. This can include steps S41-S42:

[0109] S41. The memory unit of the LSTM model obtains the memory feature corresponding to the current target feature based on the first inference result, the second inference result, the historical latent vector, and the historical memory vector corresponding to the historical memory feature.

[0110] In this embodiment of the disclosure, the memory unit can be updated based on the following computational formula:

[0111] ;

[0112] in, The memory features (feature vectors) corresponding to the target features at time t. Let t be the memory feature corresponding to the target feature at time t-1, and let ∘ represent the element-wise product (Hadamard product). These are the preset memory unit feature weight parameters. The preset implicit vector weight parameters for memory units, For the preset memory cell bias term, , , It can be acquired through learning.

[0113] S42. Obtain the latent features corresponding to the current target features based on memory features and the third reasoning result.

[0114] In this embodiment of the disclosure, the latent features (as feature vectors) at the current time t can be obtained based on the following calculation formula: ;in, The latent feature is the target feature corresponding to the target feature at the current time t.

[0115] S24. Determine the probability distribution of the current category based on the latent features.

[0116] In this embodiment of the disclosure, the probability distribution of the current category at time t can be obtained based on the following formula: Where MLP represents a two-layer feedforward neural network, x is the independent variable. These are the model learning parameters. The SELU activation function is shown in the following formula:

[0117] .

[0118] In this embodiment of the disclosure, the LSTM model improves the system's ability to abstract target features and its recognition performance by encoding and transmitting historical target features in real time.

[0119] In this embodiment of the disclosure, before using the neural network model of the scheme of this embodiment of the disclosure, the neural network can be trained first based on the collected data of the sample target to obtain the neural network model.

[0120] In the embodiments disclosed herein, such as Figure 6 As shown, the training methods for neural network models may include: S51-S52:

[0121] S51. Obtain the current sample features and historical sample features of the target sample to form training samples.

[0122] In the embodiments disclosed herein, such as Figure 7 As shown, obtaining the current sample features and historical sample features of the target sample to form training samples may include steps S61-S64:

[0123] S61. Sample the motion trajectory of the target sample based on a preset sliding window to obtain basic training samples.

[0124] In this embodiment of the disclosure, a large amount of raw data of the sample target can be collected in advance, and this raw data may include, but is not limited to, the motion trajectory of the sample target. The sample target may include, but is not limited to, vehicles.

[0125] In this embodiment of the disclosure, the raw data may include, but is not limited to, information such as the position, velocity, and reflected signal intensity of the sample target at different time points, and the motion trajectory of the sample target can be determined based on this information.

[0126] In the embodiments disclosed herein, such as Figure 8 As shown, sampling the motion trajectory of the target sample based on a preset sliding window to obtain basic training samples may include steps S71-S73:

[0127] S71. Label the category to which the sample target belongs for each motion trajectory to obtain the target motion trajectory.

[0128] In this embodiment of the disclosure, each motion trajectory can be labeled separately, and the category to which the sample target corresponding to the motion trajectory belongs can be labeled.

[0129] In this embodiment of the disclosure, the annotation information for each annotation can be defined based on one or more known categories. For example, in the motion trajectory of a vehicle, each motion trajectory can be labeled as a car, truck, bus, etc. The annotated motion trajectory serves as the target motion trajectory for subsequent model training.

[0130] S72. Based on a preset sliding window, sample each target motion trajectory according to a preset step length to obtain multiple target motion sub-trajectories.

[0131] In this embodiment of the disclosure, each target motion trajectory is a long trajectory (complete trajectory). Each target motion trajectory can be sampled by a preset sliding window with a fixed window size to obtain multiple target motion sub-trajectories with fixed lengths corresponding to the target motion trajectory.

[0132] In this embodiment, the step length of the preset sliding window during sampling can be set as needed, and is not limited in detail here. For example, each step length can be a snapshot.

[0133] In this embodiment of the disclosure, when performing sliding window sampling on each target motion trajectory, the window can be slid sequentially from the starting position of the target motion trajectory, taking one snapshot at a time to obtain a series of target motion sub-trajectories.

[0134] In the embodiments disclosed herein, such as Figure 9 The image shows a sampling example of a complete target motion trajectory with a length of 10, with the preset sliding window length set to 5. Figure 9 In the text, 1-10 represent the number of snapshots, and 0 indicates a filled snapshot. Under a filled snapshot, the feature vector of the sample target is a vector of all zeros.

[0135] S73. Use multiple target motion sub-trajectories as basic training samples.

[0136] In the embodiments of this disclosure, each target motion sub-trajectory can be regarded as an independent basic training sample.

[0137] S62. Extract features from the basic training samples to obtain the features of the current sample.

[0138] In this embodiment of the disclosure, for each basic training sample (such as a target motion sub-trajectory), the feature calculation method in the inference process is used to perform feature calculation on the point cloud data under each snapshot within the preset sliding window to obtain the features of the current sample.

[0139] S63. Obtain the historical sample features corresponding to the sample target.

[0140] In this embodiment of the disclosure, historical sample features may include sample features obtained by the neural network inferring from the sample target point cloud data of the sample target at any time before the current time.

[0141] In this embodiment of the disclosure, historical sample features may include, but are not limited to, historical sample latent features and / or historical sample memory features.

[0142] In this embodiment of the disclosure, the historical sample features may be the features of at least one cluster of historical sample target point clouds corresponding to the basic training samples of the sample target; the historical sample target point cloud is the sample target point cloud collected before the current cluster of sample target point clouds.

[0143] In this embodiment of the disclosure, the historical sample features may include, but are not limited to, the historical sample latent features output by the neural network at the previous moment (e.g., the historical sample latent feature vector) and / or the historical sample memory features (e.g., the historical sample memory feature vector).

[0144] In this embodiment of the disclosure, at the beginning of each target motion sub-trajectory, the historical sample latent feature vector and the historical sample memory feature vector are both all 0.

[0145] S64. Training samples are constructed based on the current sample features and historical sample features.

[0146] In this embodiment of the disclosure, the current sample features corresponding to the current time step of each basic training sample can be combined with the hidden features of historical samples and / or the memory features of historical samples to form a training sample, which is used as the input of the God Network (e.g., LSTM neural network).

[0147] S52. Train the preset neural network based on the training samples to obtain the neural network model.

[0148] In the embodiments disclosed herein, such as Figure 10 As shown, training a pre-defined neural network based on training samples to obtain a neural network model may include steps S81-S83:

[0149] S81. Obtain the difference between the category predicted by the neural network and the category labeled on the sample target.

[0150] In the embodiments disclosed herein, such as Figure 11 As shown, obtaining the difference between the category predicted by the neural network and the category labeled on the sample target can include steps S91-S94:

[0151] S91. Obtain the probability distribution of sample targets belonging to different categories, determined by the neural network based on the current sample features and historical sample features.

[0152] In this embodiment of the disclosure, the reasoning method is the same as that of the aforementioned neural network model. During the training process, the neural network calculates the probability distribution of the current sample target possibly belonging to different categories based on the vector of current sample features and the vector of historical sample features (e.g., the historical sample latent feature vector and / or historical sample memory feature vector from the previous moment).

[0153] S92. Determine the confidence level of the sample target belonging to each category based on the probability distribution of the sample target belonging to different categories.

[0154] In this embodiment of the disclosure, the probability distribution of the sample target belonging to different categories can be normalized. Based on the normalization result, the probability of each category in the probability distribution of the sample target belonging to different categories can be normalized to a value between 0 and 1. This value can be used as the confidence level of each category corresponding to the sample target.

[0155] S93. Determine the category of the sample target based on the confidence level of the sample target belonging to each category, as predicted by the neural network.

[0156] In this embodiment of the disclosure, the confidence level of each category can be compared with a preset confidence threshold. If the confidence level of any category is greater than or equal to the confidence threshold, the category is used as the predicted category for the sample target; or, if the confidence level of the category is less than the confidence threshold, the category is not used as the predicted category for the sample target.

[0157] S94. Determine the difference between the predicted category and the category labeled in the sample target.

[0158] S82. Adjust the network parameters of the neural network based on the differences and the preset parameter adjustment algorithm.

[0159] In this embodiment of the disclosure, the preset parameter adjustment algorithm may include, but is not limited to, the RMSProp (Root Mean Square Propagation) optimization algorithm.

[0160] In this embodiment of the disclosure, the network parameters can be further adjusted using the RMSProp optimization algorithm based on the difference between the predicted category and the category labeled on the sample target.

[0161] In this embodiment of the disclosure, the RMSProp optimization algorithm is an adaptive learning rate optimization algorithm, and its calculation process may include:

[0162] Step 1: Initialize the trainable parameters (i.e., the network parameters) in the neural network and set the hyperparameters; these neural network parameters may include, but are not limited to, weights. and bias This hyperparameter can include, but is not limited to, the learning rate. (Take 0.01), a constant used for smoothing. (Take 0.9) and the minimum value to prevent the denominator from being zero. (Take 1e-8).

[0163] Step 2: In each training iteration, based on the current input data (e.g., the difference between the predicted category and the category labeled on the sample target), the predicted output of the neural network is calculated through forward propagation. Then, the loss value is calculated based on this predicted output according to the loss function, and the gradient of each network parameter with respect to the loss function is calculated using the backpropagation algorithm based on this loss value. Let the current network parameters be... Its gradient can be denoted as ,in This is the loss function.

[0164] Step 3: Calculate the moving average of the squared gradient, introducing a cumulative variable. This is used to record the moving average of the squared gradient. The formula for calculating the moving average of the squared gradient includes...

[0165] ;

[0166] in, It is the moving average of the squared gradient calculated in the previous iteration. Initially, when hour, Initialize to 0.

[0167] Step 4: Calculate the neural network parameter update: based on the calculated gradient squared moving average. Calculate the update amount of the neural network parameters. The formula for calculating the update amount includes:

[0168] ;in, This is the learning rate.

[0169] Step 5: Use the calculated update amount of the neural network parameters Update the neural network parameters:

[0170] .

[0171] S83. Determine the final network parameters after adjusting the network parameters, and obtain the neural network model based on the neural network and the final network parameters.

[0172] In this embodiment of the disclosure, the aforementioned training process can be repeated on the entire training sample set until the model converges, that is, the prediction accuracy of the neural network on the validation dataset no longer improves significantly.

[0173] In this embodiment of the disclosure, when the model converges, the current network parameters can be obtained as the final network parameters, thereby obtaining a neural network with the final network parameters, which constitutes the neural network model.

[0174] S14. Determine the category of the target to be identified based on the current category probability distribution and the historical category probability distribution of the target to be identified.

[0175] In this embodiment of the disclosure, the historical current category probability distribution is the category probability distribution determined based on the historical target point cloud.

[0176] In the embodiments disclosed herein, such as Figure 12 As shown, determining the category of the target to be identified based on the current category probability distribution and the historical category probability distribution of the target to be identified may include steps S101-S103:

[0177] S101. Accumulate the current category probability distribution and the historical category probability distribution to obtain the target category cumulative probability distribution.

[0178] In the embodiments disclosed herein, such as Figure 13 As shown, the current category probability distribution of the target point cloud captured at the current time t can be combined with at least one historical category probability distribution. For example, the category probability distribution of the target point cloud captured at the previous time t-1 can be added together. That is, the current category probability distribution of the target point cloud captured at the current time t and the historical category probability distribution of the target point cloud captured at the previous time t-1 can be used together for subsequent category recognition.

[0179] In this embodiment of the disclosure, the current category probability distribution and the historical category probability distribution are combined to obtain the target category cumulative probability distribution.

[0180] S102. Determine the confidence level of each category corresponding to the target to be identified based on the cumulative probability distribution of the target categories.

[0181] In this embodiment of the disclosure, the cumulative probability distribution of the target category can be normalized. For example, the safe-softmax function can be used for normalization. The safe-softmax function is a commonly used normalization function that normalizes a vector containing K (K is a positive integer) real numbers. Mapped to a probability vector Each element ( All elements are between 0 and 1, and the sum of all elements is 1. The formula for calculating the safe-softmax function is as follows:

[0182] ;

[0183] in, Let j represent the j-th element, where j is a positive integer, and 1 ≤ j ≤ K.

[0184] In this embodiment of the disclosure, based on the normalization processing of the cumulative probability distribution of the target categories, the probability of each category in the cumulative probability distribution of the target categories is normalized to a value between 0 and 1, which can be used as the confidence level of each category corresponding to the target to be identified.

[0185] S103. Determine the category of the target to be identified based on the confidence level of each category.

[0186] In the embodiments disclosed herein, such as Figure 14 As shown, determining the category of the target to be identified based on the confidence level of each category may include steps S111-S112:

[0187] S111. Compare the confidence level of each category with the preset confidence level threshold.

[0188] In this embodiment of the disclosure, the confidence threshold can be set in advance according to the requirements, and the specific value of the confidence threshold is not limited here.

[0189] S112. In response to any first category having a confidence level greater than or equal to a confidence threshold, the first category is used as the predicted category of the target to be identified; or, in response to the first category having a confidence level less than a confidence threshold, the first category is not used as the predicted category of the target to be identified.

[0190] In this embodiment of the disclosure, if the confidence level of all categories is less than the confidence level threshold, the target is considered to be another unknown category. If the confidence level of multiple categories is greater than or equal to the confidence level threshold, the category with the highest confidence level can be used as the predicted category of the target to be identified.

[0191] In this embodiment of the disclosure, the type of the target to be identified is not only determined by the current category probability distribution at the current moment, but also by accumulating the historical category probability distribution at historical moments, effectively integrating the information of the target to be identified at different time points, and improving the decision-making accuracy and robustness of the system.

[0192] The present disclosure includes at least the following advantages:

[0193] 1. Efficient target feature extraction:

[0194] a) Technical problems to be solved: Target feature extraction methods in related technologies suffer from high computational overhead, poor versatility, and limited flexibility. For example, micro-Doppler-based feature extraction methods are computationally complex, while single-snapshot point cloud recognition methods based on traditional machine learning involve cumbersome manual feature engineering and struggle to extract high-order features.

[0195] b) The technical feature adopted in the embodiments of this disclosure: a feature extraction method for physical measurements based on trajectory point clouds.

[0196] c) Beneficial Effects: It can comprehensively reflect multiple motion characteristics and behavioral patterns of the target under low computational resource requirements. Compared with related technologies, it greatly reduces computational overhead and improves the efficiency of target recognition data preprocessing. Because it directly utilizes the target point cloud and its physical measurements for feature extraction, it avoids complex time-frequency analysis (such as STFT based on micro-Doppler methods) and cumbersome manual feature engineering operations, enabling the system to process data more quickly and provide a more accurate and comprehensive feature foundation for subsequent target recognition.

[0197] 2. Powerful target feature modeling capabilities

[0198] a) Technical problem to be solved: Traditional feature modeling methods have difficulty capturing the long-term dependencies of target features and have poor adaptability to complex scenarios and target behaviors.

[0199] b) Technical features adopted in the embodiments of this disclosure: using a multi-layer LSTM model as a feature modeling tool.

[0200] c) Beneficial Effects: The LSTM model improves the system's ability to abstract target features and its recognition performance through real-time automatic encoding and transmission of historical target features. Compared to traditional methods, it can better capture the long-term dependencies of target features, making the system's modeling of target features more accurate and comprehensive. This helps the system better adapt to various complex scenarios and target behaviors. For example, in dynamic scenarios, it can more accurately identify high-order dynamic information such as changes in target trajectory and speed trends, thereby improving the accuracy of target recognition.

[0201] 3. Excellent technical architecture scalability

[0202] a) Technical problems to be solved: When faced with new targets or complex environmental changes, related technologies often struggle to be effectively upgraded and expanded.

[0203] b) The technical features adopted in the embodiments of this disclosure are: good scalability in technical architecture, easy to integrate with advanced feature extraction technology, for example, it can be easily extended to PointNet feature extraction, and can input disordered point cloud data into a point cloud feature extraction network before proceeding to subsequent steps.

[0204] c) Beneficial Effects: This scalability allows the solutions of this disclosure to better adapt to future technological developments and changes in application scenarios. When facing new targets or complex environments, new feature extraction techniques can be easily introduced to further improve system performance. Compared with related technologies, it is not helpless in the face of new challenges due to limitations in its technical architecture, providing the possibility for continuous system upgrades and optimization.

[0205] 4. Improve decision-making accuracy and robustness

[0206] a) Technical problem to be solved: In actual target recognition scenarios, due to the complex and ever-changing environment, the measured target features may not be accurate, which leads to poor decision-making accuracy and robustness in target recognition.

[0207] b) The technical feature adopted in the embodiments of this disclosure is that a probability accumulation method is introduced based on the probability distribution of the LSTM model output at each time step.

[0208] c) Beneficial Effects: This method effectively integrates information about the target at different points in time. Even if the measured features are not accurate at certain moments, by accumulating historical probability distributions (accumulating over time), the system can more comprehensively consider the target's behavioral trends, reducing the impact of inaccurate single measurements, thereby improving the system's decision-making accuracy and robustness. This enables the system to perform target identification more stably and reliably in complex environments. Compared to related technologies, this approach based on comprehensive consideration of historical trajectories provides a more solid foundation for decision-making and improves the system's stability.

[0209] This disclosure also provides an electronic device 100, such as... Figure 15 As shown, it includes:

[0210] One or more processors 101;

[0211] The memory 102 stores one or more programs that, when executed by the one or more processors, enable the one or more processors 101 to implement the category recognition method.

[0212] One or more input / output (I / O) interfaces 103 are connected between the processor 101 and the memory 102 and configured to enable information interaction between the processor 101 and the memory 102.

[0213] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the category recognition method described above.

[0214] This disclosure also provides a computer program product, which includes a computer program that, when executed by a processor, implements the category recognition method.

[0215] In the embodiments disclosed herein, any of the aforementioned category identification methods are applicable to the embodiments of the electronic device, storage medium, and program product, and will not be described in detail here.

[0216] Those skilled in the art will understand that all or some of the functional modules / units disclosed above can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0217] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be executed by several physical components working together.

[0218] Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit (CPU), digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technique for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory (FLASH) or other disk storage; read-only optical disc (CD-ROM), digital versatile disc (DVD) or other optical disc storage; magnetic cartridges, magnetic tapes, disk storage or other magnetic storage; and any other media that can be used to store desired information and can be accessed by a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0219] This disclosure has disclosed exemplary embodiments, and although specific terminology has been used, it is for general illustrative purposes only and should not be construed as limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of this disclosure as set forth by the appended claims.

Claims

1. A category recognition method, comprising: Obtain the target point cloud corresponding to the target to be identified; Extract the current target features based on the physical measurements of the target point cloud; The current category probability distribution is determined based on the current target features and the historical target features of the target to be identified; The category of the target to be identified is determined based on the current category probability distribution and the historical category probability distribution of the target to be identified.

2. The category identification method according to claim 1, wherein, The physical measurements of the target point cloud include: the spatial coordinates and / or echo power of each point in the target point cloud; The current target features include at least one of the following: the number of points in the target point cloud, the point cloud diameter, and the standard deviation of the echo power.

3. The category identification method according to claim 2, wherein, The extraction of current target features based on physical measurements of the target point cloud includes at least one of the following: Count the number of points in the target point cloud; The point cloud diameter of the target point cloud is determined based on the spatial coordinates of each point; and, The standard deviation of the echo power is calculated based on the echo power to obtain the standard deviation of the echo power.

4. The category recognition method according to claim 1, after obtaining the target point cloud corresponding to the target to be identified, the method further includes: Obtain the unordered point cloud data of the target point cloud; The disordered point cloud data is input into a preset point cloud neural network model; Based on the point cloud neural network model, inference is performed on the disordered point cloud data to extract the current target features.

5. The category identification method according to claim 1, wherein, The current target feature is the feature of the current cluster of target point clouds corresponding to the target to be identified, and the historical target feature is the feature of at least one cluster of historical target point clouds corresponding to the target to be identified; the historical target point cloud is a target point cloud collected before the current cluster of target point clouds. The step of determining the current category probability distribution based on the current target features and the historical target features of the target to be identified includes: Based on a preset neural network model, reasoning is performed on the current target features and the historical target features to obtain the probability distribution of the current category; The neural network model is capable of processing time-series data and capturing the temporal dependencies between the time-series data.

6. The category identification method according to claim 5, wherein, The historical target features include historical latent features and / or historical memory features obtained by the neural network model through reasoning on the historical target point cloud.

7. The category identification method according to claim 6, wherein, The neural network model includes a Long Short-Term Memory (LSTM) model; the method based on the preset neural network model, which infers the current target features and the historical target features to obtain the current category probability distribution, includes: The current target features are preprocessed to obtain a normalized target feature vector; The normalized target feature vectors are input into the LSTM model, and the LSTM model obtains the inference results based on the normalized target feature vectors and the historical latent vectors corresponding to the historical latent features. The LSTM model obtains the latent features corresponding to the current target features based on the inference results, the historical latent vectors, and the historical memory vectors corresponding to the historical memory features. The probability distribution of the current category is determined based on the hidden features.

8. The category identification method according to claim 7, wherein, The reasoning results include a first reasoning result, a second reasoning result, and a third reasoning result; The step of inputting the normalized target feature vector into the LSTM model, and having the LSTM model obtain the inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent features, includes: The normalized target feature vector is input into the input gate of the LSTM model, and the input gate obtains the first inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent feature. The normalized target feature vector is input into the forget gate of the LSTM model, and the forget gate obtains the second inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent feature. The normalized target feature vector is input into the output gate of the LSTM model, and the output gate obtains the third inference result based on the normalized target feature vector and the historical latent vector corresponding to the historical latent feature.

9. The category identification method according to claim 8, wherein, The LSTM model obtains the latent features corresponding to the current target feature based on the inference result, the historical latent vector, and the historical memory vector corresponding to the historical memory feature, including: The memory unit of the LSTM model obtains the memory feature corresponding to the current target feature based on the first inference result, the second inference result, the historical latent vector, and the historical memory vector corresponding to the historical memory feature; The latent features corresponding to the current target features are obtained based on the memory features and the third reasoning result.

10. The category identification method according to claim 1, wherein, The training methods for the neural network model include: Obtain the current and historical sample features of the target sample to form training samples; The neural network model is obtained by training the preset neural network based on the training samples.

11. The category identification method according to claim 10, wherein, The current and historical sample features of the target sample are obtained to form training samples, including: The motion trajectory of the target sample is sampled based on a preset sliding window to obtain basic training samples; Feature extraction is performed on the basic training samples to obtain the features of the current sample; Obtain the historical sample features corresponding to the target sample; The training samples are constructed based on the current sample features and the historical sample features.

12. The category identification method according to claim 11, wherein, The step of sampling the motion trajectory of the sample target based on a preset sliding window to obtain basic training samples includes: Label each of the motion trajectories to which the sample target belongs, and obtain the target motion trajectory; Based on the preset sliding window, each target motion trajectory is sampled according to a preset step length to obtain multiple target motion sub-trajectories; The multiple target motion sub-trajectories are used as the basic training samples.

13. The category identification method according to claim 11, wherein, The historical sample features include historical sample latent features and / or historical sample memory features; The historical sample features are the features of at least one cluster of historical sample target point clouds corresponding to the sample target; the historical sample target point cloud is the sample target point cloud collected before the current cluster of sample target point cloud.

14. The category identification method according to claim 10, wherein, The step of training a preset neural network based on the training samples to obtain the neural network model includes: The difference between the category of the sample target predicted by the neural network and the category labeled on the sample target is obtained; The network parameters of the neural network are adjusted based on the aforementioned differences and a preset parameter adjustment algorithm. Determine the final network parameters after adjusting the network parameters, and obtain the neural network model based on the neural network and the final network parameters.

15. The category identification method according to claim 14, wherein, The training samples include: current sample features and historical sample features of the target sample; The step of obtaining the difference between the category predicted by the neural network and the category labeled on the sample target includes: Obtain the probability distribution of the sample target belonging to different categories, determined by the neural network based on the current sample features and the historical sample features; The confidence level of the sample target belonging to each category is determined based on the probability distribution of the sample target belonging to different categories; The category of the sample target predicted by the neural network is determined based on the confidence level of the sample target belonging to each category; Determine the difference between the predicted category and the category labeled on the sample target.

16. The category identification method according to claim 1, wherein, The historical category probability distribution is determined based on the historical target point cloud; Determining the category of the target to be identified based on the current category probability distribution and the historical category probability distribution of the target to be identified includes: The current category probability distribution and the historical category probability distribution are summed to obtain the target category cumulative probability distribution; The confidence level of each category corresponding to the target to be identified is determined based on the cumulative probability distribution of the target categories; The category of the target to be identified is determined based on the confidence level of each category.

17. The category identification method according to claim 16, wherein, Determining the category of the target to be identified based on the confidence level of each category includes: The confidence level of each category is compared with a preset confidence threshold. In response to any first category having a confidence level greater than or equal to the confidence threshold, the first category is used as the predicted category of the target to be identified; or, in response to the first category having a confidence level less than the confidence threshold, the first category is not used as the predicted category of the target to be identified.

18. An electronic device comprising: One or more processors; A memory having stored one or more programs thereon, which, when executed by one or more processors, cause the one or more processors to implement the category recognition method according to any one of claims 1-17; One or more input / output (I / O) interfaces are connected between the processor and the memory and configured to enable information exchange between the processor and the memory.

19. A computer program product comprising a computer program that, when executed by a processor, implements the category recognition method according to any one of claims 1-17.

Images