Recognition model training methods, recognition methods, computer equipment, and computer-readable storage media

Abstract

This application discloses a method for training a recognition model, a recognition method, a computer device, and a computer-readable storage medium. The training method includes the following steps: controlling a single-photon lidar to emit photon pulses towards a target scene and receiving one-dimensional single-photon echoes reflected from the target scene, where the target scene includes at least one object to be recognized; constructing a training sample set based on the one-dimensional single-photon echoes; constructing an initial model and inputting the training sample set into the initial model for iterative training; validating the initial model after each iteration, marking the validated initial model as the recognition model, and outputting it. The application method involves inputting the echo signal to be recognized into the trained recognition model, and obtaining the recognition result based on the processing of the recognition model. The recognition result includes the pose and type of the object in the target scene. Therefore, this application can achieve high-precision recognition of long-distance targets.

Description

Technical Field

[0001] This application belongs to the field of lidar technology, and specifically relates to a recognition model training method, a single-photon lidar image-free recognition method, a computer device, and a computer-readable storage medium. Background Technology

[0002] Target recognition technology has significant application value in fields such as autonomous driving, aerial surveillance, and long-range detection. Accurately acquiring target category and attitude information at distances of hundreds of meters or even longer is a core requirement in current engineering systems. As application scenarios become increasingly complex, higher demands are placed on the reliable identification capabilities of long-range targets. Single-photon lidar, with its single-photon-level detection sensitivity and picosecond-level temporal resolution, has become an important technological means for long-range target detection.

[0003] A data-driven, image-free single-photon lidar identification method is proposed in the existing technology. This method illuminates the target with pulsed laser light, uses a single-point, non-scanning single-photon avalanche diode to record the flight time of backscattered photons, forming a one-dimensional single-photon echo reflecting the target's depth distribution, and combines a one-dimensional convolutional neural network to directly extract target features from the time data, achieving target category and attitude recognition without the need for images.

[0004] However, such methods heavily rely on a large number of high-quality, accurately labeled training samples. In real-world long-range detection environments, due to the diversity of target types, complex pose variations, and high on-site acquisition costs, it is difficult to obtain a sufficiently large and accurately labeled training dataset. Insufficient training data will prevent the neural network model from fully learning target features, thus significantly limiting recognition performance and making it difficult to meet practical engineering needs.

[0005] Therefore, there is an urgent need for a technical solution and device that can achieve image-free recognition for long-range single-photon lidar without relying on large-scale training datasets. The preceding description is intended to provide general background information and does not necessarily constitute existing technology. Summary of the Invention

[0006] Based on this, it is necessary to propose a recognition model training method, a single-photon lidar image-free recognition method, computer equipment, and computer-readable storage medium to address the above problems, which can achieve high-precision recognition of long-distance targets.

[0007] The technical problem solved by this application is achieved by the following technical solution:

[0008] This application provides a method for training a recognition model, comprising the following steps: controlling a single-photon lidar to emit photon pulses toward a target scene and receiving a one-dimensional single-photon echo reflected from the target scene, wherein the target scene includes at least one object to be recognized; constructing a training sample set based on the one-dimensional single-photon echo; constructing an initial model and inputting the training sample set into the initial model for iterative training; validating the initial model after each iteration, marking the validated initial model as the recognition model and outputting it.

[0009] In an optional embodiment of this application, controlling a single-photon lidar to emit photon pulses toward a target scene and receive a one-dimensional single-photon echo reflected from the target scene includes: setting the emission parameters of the single-photon lidar, the emission parameters including amplitude and frequency; determining an emission pulse signal function based on the emission parameters, the emission pulse signal function being used to control the single-photon lidar to emit photon pulses; acquiring reception parameters, and determining the average number of detected photons within one pulse period based on the reception parameters and the emission pulse signal function; calculating the photon arrival probability distribution within each time slot based on the average number of detected photons; and accumulating the arrival probability distributions of all pulse periods to generate a one-dimensional single-photon echo.

[0010] In an optional embodiment of this application, the photon arrival probability distribution in each time slot is calculated based on the average number of detected photons, including: modeling the average number of detected photons using Poisson statistics to obtain the detection probability, which describes the probability of detecting a photon in the nth time slot; determining the photon arrival probability in each time slot based on the detection probability, denoted as the arrival probability; obtaining a preset probability decision threshold, filtering all arrival probabilities, and marking the time slots corresponding to arrival probabilities greater than the probability decision threshold as valid time slots; and summing the arrival probabilities corresponding to all valid time slots within a period to obtain the photon arrival probability distribution for the corresponding period.

[0011] In an optional embodiment of this application, constructing a training sample set based on one-dimensional single-photon echoes includes: obtaining preset screening conditions and screening one-dimensional single-photon echoes; summarizing one-dimensional single-photon echoes that meet the screening conditions; setting them as source domain training sets according to the identification object; setting labels; setting one-dimensional single-photon echoes with labels as source domain training sets; summarizing one-dimensional single-photon echoes that do not meet the screening conditions and setting them as target domain training sets; and constructing a training sample set based on the source domain training set and the target domain training set.

[0012] In an optional embodiment of this application, the training sample set includes a source domain training set and a target domain training set; constructing an initial model involves inputting the training sample set into the initial model for iterative training, including: constructing an initial model, which includes a feature extractor, a classifier, and a domain discriminator; the feature extractor uses the training sample set as input to extract feature vectors, which are then used as input to the classifier and the domain discriminator; the classifier uses the feature vectors from the source domain training set as input to process and output predicted labels; the domain discriminator uses the feature vectors from both the source and target domain training sets as input to process and output discrimination results, which indicate whether the feature vectors come from the source or target domain training set; the training sample set is input into the initial model for processing to obtain feature vectors, predicted labels, and discrimination results; a loss function is constructed, and the feature vectors, predicted labels, and discrimination results are used as input to the loss function to calculate training parameters; the initial model is then updated using the training parameters.

[0013] In an optional embodiment of this application, constructing a loss function includes: obtaining the first total number of samples and the true labels in the source domain training set to construct a classification loss function; obtaining the second total number of samples in the target domain training set; constructing a domain discrimination loss function based on the second total number of samples and the first total number of samples; constructing a feature distribution alignment loss function based on the first total number of samples and the second total number of samples; and using the classification loss function, the domain discrimination loss function, and the feature distribution alignment loss function for network training.

[0014] In an optional embodiment of this application, the initial model is validated after each iteration, including: after each iteration, obtaining the predicted labels output by the initial model processing the training sample set; obtaining the real labels within the training sample set; obtaining the classification accuracy based on the predicted labels and the real labels; determining that the initial model has passed validation when the classification accuracy meets a preset validation condition; and / or, after each iteration, obtaining the current iteration number of the initial model; determining that the initial model has passed validation when the current iteration number meets a preset exit condition.

[0015] This application also provides a single-photon lidar image-free recognition method, including the following steps: acquiring echo data collected by the single-photon lidar for a target scene; inputting the echo data into a recognition model, which is obtained according to the training method provided above; processing the echo data through the recognition model to obtain a recognition result, which includes the pose and type of objects in the target scene.

[0016] This application also provides a computer device including a processor and a memory: the processor is used to execute a computer program stored in the memory to implement the method as described above.

[0017] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described above.

[0018] The embodiments of this application have the following beneficial effects:

[0019] Compared to existing data-driven single-photon radar recognition methods, this application employs a domain-adversarial transfer learning-based single-photon radar recognition technology. It utilizes readily available near-range labeled data as the source domain and combines it with long-range unlabeled data as the target domain for training. This eliminates the need to collect large amounts of labeled data in long-range scenarios, significantly reducing data acquisition costs and labeling workload. Furthermore, the training process employs multiple loss functions compared to a single loss function, overcoming the limitations of a single loss function's limited scope and adaptability, thus achieving comprehensive optimization of model training. This effectively reduces recognition errors caused by distribution differences between the source and target domains, ensuring the model maintains high efficiency and reliable recognition performance even in long-range unlabeled scenarios. Compared to the limitations of traditional multilayer perceptrons and basic convolutional networks in fully mining single-photon echo features, this application improves the model architecture. The feature extractor utilizes a design that downsamples and compresses features through multilayer encoding paths and upsamples and restores resolution through multilayer decoding paths. This effectively enhances the feature extraction capability for weak long-range echo signals, improving target recognition accuracy at long distances.

[0020] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it according to the contents of the specification, and to make the above and other objects, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit this application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a recognition model training method provided in one embodiment.

[0023] Figure 2 This is a schematic diagram illustrating an application scenario for collecting one-dimensional single-photon echoes, as provided in one embodiment.

[0024] Figure 3 A schematic diagram of the network architecture of an initial model provided for one embodiment.

[0025] Figure 4 This is a flowchart illustrating a single-photon lidar image-free recognition method according to one embodiment.

[0026] Figure 5 This is a schematic diagram illustrating the experimental results of an existing identification method provided in one embodiment.

[0027] Figure 6 This is a schematic diagram illustrating the experimental results of the method provided in this application as an embodiment.

[0028] Figure 7 This is a schematic block diagram of the structure of a computer device provided in one embodiment. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0030] In existing technologies, a pulsed laser is used to illuminate the target with floodlight. A single-point, non-scanning single-photon avalanche diode is used to record the flight time of backscattered photons, forming a one-dimensional single-photon echo that reflects the depth distribution of the target. This echo is then combined with a one-dimensional convolutional neural network to directly extract target features from the time data, enabling target category and pose recognition without the need for images.

[0031] However, such methods heavily rely on a large number of high-quality, accurately labeled training samples. In real-world long-range detection environments, due to the diversity of target types, complex pose changes, and high on-site acquisition costs, it is difficult to obtain a sufficiently large and accurately labeled training dataset. Insufficient training data will prevent the neural network model from fully learning target features, thus significantly limiting recognition performance and making it difficult to meet practical engineering needs. To alleviate the performance bottleneck caused by the scarcity of training data, this application proposes a method for training and applying a recognition model for long-range single-photon lidar without relying on a large-scale training dataset. For a clear description of the recognition model training method provided in this embodiment, please refer to [reference needed]. Figures 1-6 This includes steps S110 to S140.

[0032] Step S110: Control the single-photon lidar to emit photon pulses toward the target scene and receive one-dimensional single-photon echoes reflected back from the target scene, the target scene including at least one object to be identified.

[0033] In one embodiment, for applications where single-photon lidar collects one-dimensional single-photon echoes, reference can be made to... Figure 2 As shown, the control pulse emission source emits pulses towards the target scene and triggers a start signal to the timing module. After the pulse signal is reflected by the target, it is received by the single-photon detector, which sends a stop signal to the timing module to record the arrival time of the photon. Through the accumulation of multiple pulses, the system finally obtains a one-dimensional single-photon echo containing information about the target scene. Figure 2 The TCSPC module (Time-Correlated Single-Photon Counting) is an electronic timing system used to record photon arrival times and generate time histograms. The display and control terminal is used to display the histogram and for subsequent calculations.

[0034] In one embodiment, step S110: controlling the single-photon lidar to emit photon pulses toward the target scene and receiving one-dimensional single-photon echoes reflected from the target scene, includes: setting the emission parameters of the single-photon lidar, the emission parameters including amplitude and frequency; determining the emission pulse signal function according to the emission parameters, the emission pulse signal function being used to control the single-photon lidar to emit photon pulses; acquiring the receiving parameters, determining the average number of detected photons within one pulse period according to the receiving parameters and the emission pulse signal function; calculating the photon arrival probability distribution within each time slot according to the average number of detected photons; accumulating the arrival probability distributions of all pulse periods to generate a one-dimensional single-photon echo.

[0035] In one embodiment, to facilitate the capture and generation of one-dimensional single-photon echoes, the emitted pulse signal needs to be precisely controlled by setting appropriate pulse width and repetition frequency, i.e., setting the emission parameters of the single-photon lidar. These emission parameters include amplitude and frequency to match the characteristics of targets with different distances and reflectivities. The emission pulse signal function can be found in the following formula.

[0036] (1)

[0037] The transmit pulse signal function is used to control the single-photon lidar to emit photon pulses. Among them, For amplitude, For frequency, For time. The signal envelope can be represented as follows.

[0038] (2)

[0039] In the above formula, This is a constant related to pulse width, and is a preset value. (Reference) Figure 2The single-photon lidar, controlled by a function of the emitted pulse signal, emits photon pulses toward the target scene, while simultaneously activating a time-correlated single-photon counting module for time sampling. The target scene includes at least one object to be identified. Figure 2 As shown, the object to be identified can be an aircraft. In the target scene, referring to reality, the objects to be identified are usually arranged in a staggered manner, with varying distances between them and different orientations of their reflective surfaces. Real-world identification requires detecting the type, distance, and pose of the objects. After the photon pulse propagates to the target scene, it is scattered and reflected by each object to be identified, and returns to the single-photon detector for detection. After reflection from the target, according to the radar equation, the power of the echo pulse rapidly attenuates, and the single-photon detector will receive a certain number of photons within a specified time. The average number of photons detected within one pulse cycle can be determined based on the received parameters and the transmitted pulse signal function; the calculation method is as follows.

[0040] (3)

[0041] In the above formula, This indicates the detection efficiency of a single-photon detector. The signal photon frequency is included in the receiving parameters. This represents the distance to each identified object in the target scene, where S represents the target plane, which can be directly determined from the received photons. It also includes several constants, including Planck's constant. speed of light and in Number of noise photons received within a time period , n This indicates the sequence number of the time slot. It can be understood that within a pulse period, time can be divided into several time slots for time sampling, with each time slot having a preset width.

[0042] Based on the average number of detected photons, the arrival probability distribution of photons in each time slot is calculated; the arrival probability distribution of all pulse periods is accumulated to generate a one-dimensional single-photon echo.

[0043] In one embodiment, the photon arrival probability distribution within each time slot is calculated based on the average number of detected photons, including: modeling the average number of detected photons using Poisson statistics to obtain the detection probability, which describes the probability of detecting a photon in the nth time slot; determining the photon arrival probability within each time slot based on the detection probability, denoted as the arrival probability; obtaining a preset probability decision threshold, filtering all arrival probabilities, and marking the time slots corresponding to arrival probabilities greater than the probability decision threshold as valid time slots; and summing the arrival probabilities corresponding to all valid time slots within a period to obtain the photon arrival probability distribution for the corresponding period.

[0044] In one embodiment, the essence of a one-dimensional single-photon echo is the detection probability distribution of echo photons in each detection time slot. This distribution reflects the spatial layout characteristics of each identified object in the target scene in the distance dimension. The probability of detecting 0 photons in the nth time slot in a single pulse echo is... The probability of detecting one or more photons is given by the following formula.

[0045] (4)

[0046] Under the multi-thermal-optical field approximation, the photon detection process of echo signals from general rough targets can be completely described using Poisson statistical distribution. Therefore, Poisson statistics are used to model the average number of detected photons to obtain the detection probability, which describes the probability of detecting a photon in the nth time slot. The detection probability can be expressed as follows.

[0047] (5)

[0048] Considering the dead time of a single-photon detector, the probability of detecting a photon for the first time in the nth time slot can be obtained by combining equations (4) and (5), denoted as the arrival probability. The arrival probability can be expressed as:

[0049] (6)

[0050] The probability of the first detection event in each time slot can be calculated from the received echo signal using formula (6), which can be expressed as the following formula.

[0051] (7)

[0052] In the above formula, This represents the total number of time slots within a pulse period. For each time slot, a preset probability decision threshold can be obtained. The arrival probability of each time slot is filtered. This probability decision threshold... Following a uniform distribution between 0 and 1, the arrival probabilities are compared with a probability decision threshold. Time slots with arrival probabilities greater than the threshold are considered valid time slots, while others are considered to have no target response. For valid time slots, the number of photons detected is recorded. The arrival probabilities of all valid time slots within a period are summarized to obtain the photon arrival probability distribution for that period. After all time slots within a period have been processed, the echo data acquisition and processing for the next pulse period begins. The arrival probability distributions of all pulse periods are accumulated to generate a one-dimensional single-photon echo. The one-dimensional single-photon echo is presented as a statistical time histogram of target scene information.

[0053] In one embodiment, step S120: construct a training sample set based on one-dimensional single-photon echoes.

[0054] In one embodiment, step S120: constructing a training sample set based on one-dimensional single-photon echoes includes: obtaining preset screening conditions and screening one-dimensional single-photon echoes; summarizing one-dimensional single-photon echoes that meet the screening conditions; setting them as source domain training sets according to the identification object; setting labels; setting one-dimensional single-photon echoes with labels as source domain training sets; summarizing one-dimensional single-photon echoes that do not meet the screening conditions and setting them as target domain training sets; and constructing a training sample set based on the source domain training set and the target domain training set.

[0055] In one embodiment, as described above, the one-dimensional single-photon echo is actually obtained by emitting laser pulses at the target object and detecting its echo signal. The echo data acquired in this process contains information such as the distance, shape, and orientation of the target object. It is understandable that as the distance increases, the attenuation of the photon echo becomes more significant, and the signal-to-noise ratio gradually decreases, leading to a decrease in the detection probability of distant targets. This results in insufficient samples corresponding to distant targets during the training of the recognition model. Insufficient training data will prevent the neural network model from fully learning the target features, thus significantly limiting the recognition performance and making it difficult to meet practical engineering needs.

[0056] To alleviate the decline in model generalization ability caused by the scarcity of distant target samples, and to learn the deep feature distribution patterns of distant samples, this application classifies the training dataset during training. Preset filtering conditions are obtained to filter one-dimensional single-photon echoes. Specifically, these filtering conditions can be the distance range between the one-dimensional single-photon echo and the corresponding object to be identified. For example, samples in the echo signal with a distance greater than a preset threshold are classified as distant target samples, i.e., the target domain training set; those less than or equal to the threshold are classified as near-distance target samples, i.e., the source domain training set.

[0057] To this end, the training set is divided into a source domain training set and a target domain training set, with aircraft seats as an example of the object to be identified. The source domain training set includes three different aircraft, generating a large number (e.g., 2000 images) of each by changing their attitude and angle, and using single-photon radar to generate corresponding one-dimensional single-photon echoes. Based on the one-dimensional single-photon echoes, different aircraft in the object to be identified are standardized to determine corresponding labels. These labels can represent information such as the aircraft's model, category, and attitude, thus forming more pairs (e.g., 6000 pairs) of paired data between different aircraft labels and one-dimensional single-photon echoes. This paired data constitutes the samples in the source domain training set. The target domain training set only contains one-dimensional single-photon echoes, i.e., collected from long-distance echo samples that are difficult to obtain in actual detection, without corresponding labels. Furthermore, for the samples in the training sample set, the data can be expanded by applying methods such as changing the target's attitude and position to the original one-dimensional single-photon echo data, thereby increasing the number of samples.

[0058] Step S130: Construct an initial model by inputting the training sample set into the initial model for iterative training.

[0059] In one embodiment, the training sample set includes a source domain training set and a target domain training set; constructing an initial model involves inputting the training sample set into the initial model for iterative training, including: constructing an initial model, which includes a feature extractor, a classifier, and a domain discriminator; the feature extractor takes the training sample set as input to extract feature vectors, which are then used as input to the classifier and the domain discriminator; the classifier takes the feature vectors from the source domain training set as input to process and output predicted labels; the domain discriminator takes the feature vectors from both the source and target domain training sets as input to process and output a discrimination result, which indicates whether the feature vector comes from the source or target domain training set; the training sample set is input into the initial model for processing to obtain feature vectors, predicted labels, and discrimination results; a loss function is constructed, and the feature vectors, predicted labels, and discrimination results are used as input to the loss function to calculate training parameters; the initial model is then updated using the training parameters.

[0060] In one implementation, an initial model is constructed, which employs a domain adversarial neural network architecture, including a feature extractor, a classifier, and a domain discriminator. The model architecture of the initial model can be found in [reference needed]. Figure 3 As shown. The feature extractor, as input to the initial model, is used to receive and process samples from the training sample set.

[0061] The feature extractor can employ a one-dimensional UNet architecture, including an encoding path and a decoding path. The encoding path progressively compresses and extracts multi-scale features through 5 downsampling layers, increasing the number of channels from 1 to 512. The decoding path restores the feature resolution through 4 upsampling layers and fuses features from corresponding layers of the encoding path via skip connections. Finally, global average pooling compresses the feature map into a fixed-dimensional feature vector. The feature extractor is connected to the classifier and domain discriminator, respectively, and the processed feature vectors are input into the classifier and domain discriminator for further processing.

[0062] The classifier takes the feature vectors from the source domain training set as input. Its architecture can employ a three-layer fully connected network, with each layer undergoing non-linear transformations using the ReLU activation function and Dropout regularization. The final output is a probability distribution of the predicted labels. These predicted labels indicate the category, pose, etc., of the object to which the sample belongs.

[0063] The domain discriminator takes feature vectors from the source and target domain training sets as input. It can use the same three-layer fully connected structure as the classifier, but its output dimension is 2. The output of the domain discriminator represents the probability distribution of whether a sample originates from the source or target domain, and is used to achieve inter-domain feature alignment.

[0064] In one embodiment, constructing a loss function includes: obtaining the first total number of samples and the true labels in the source domain training set to construct a classification loss function; obtaining the second total number of samples in the target domain training set; constructing a domain discrimination loss function based on the second total number of samples and the first total number of samples; constructing a feature distribution alignment loss function based on the first total number of samples and the second total number of samples; and weighted summing the classification loss function, the domain discrimination loss function, and the feature distribution alignment loss function to obtain the loss function.

[0065] In one embodiment, for long-range target recognition applications using single-photon radar, employing multiple loss functions, compared to a single loss function, can overcome the limitations of a single loss function, such as its one-sided target coverage and weak adaptability, thus achieving comprehensive optimization of model training. Therefore, the constructed loss function consists of a classification loss function, a domain discriminant loss function, and a feature distribution alignment loss function.

[0066] The classification loss function is constructed by obtaining the total number of samples and the true labels in the source domain training set. This function trains the feature extractor and classifier using the labeled data from the source domain, ensuring the model learns discriminative features that can distinguish different categories. The classification loss only applies to the source domain data, as the target domain data is unlabeled. Its mathematical expression can be found in the following formula.

[0067] (8)

[0068] In the above formula, This represents the total number of samples in the first sample set, which is also the number of samples in the source domain training set. Represented as the first i Each sample belongs to category C The true label, Represented as the prediction of the first i Each sample belongs to category C The probability of.

[0069] Obtain the second total number of samples in the target domain training set; construct a domain discriminant loss function based on the second total number of samples and the first total number of samples. This loss is used to train the feature extractor and the domain discriminator. The domain discriminant loss function trains the domain discriminator to identify features from the source domain (labeled 0) and features from the target domain (labeled 1). Through a gradient reversal layer, this loss acts inversely on the feature extractor and the domain discriminator, forcing the feature extractor to generate features that the domain discriminator cannot distinguish as belonging to the source or target domain, thereby eliminating source domain-specific feature patterns and enabling the feature extractor to learn domain-invariant features. Its mathematical expression can be found in the following formula.

[0070] (9)

[0071] In the above formula, This represents the total number of the second set of samples, which is also the number of samples in the training set for the target domain. For real labels, For predicted labels.

[0072] A feature distribution alignment loss function is constructed based on the total number of the first sample and the total number of the second sample, and used to train the feature extractor. The feature distribution alignment loss function directly measures the distance between the feature distributions of the source and target domains through the maximum mean difference of multiple kernels, explicitly aligning the feature distributions of the two domains to the same space by minimizing this distance. Using multiple Gaussian kernels of different scales allows for matching distribution features at different granularities, and the kernel weights can be learned to allow the network to adaptively determine the importance of each scale. Its mathematical expression can be shown in the following equation.

[0073] (10)

[0074] In the above formula, This is represented as the feature vector extracted by the feature extractor. Specifically, For the first i Feature vectors extracted from samples in the source domain; For the first j Feature vectors extracted from samples in the source domain; For the first i Feature vectors extracted from samples in the target domain; For the first j Feature vectors extracted from samples in the target domain. The feature distribution alignment loss function calculates the loss and applies only to the feature extractor, ensuring its learning domain remains unchanged in the feature representation. The function... The calculation method is shown in the following formula.

[0075] (11)

[0076] In the above formula, This is a preset constant.

[0077] Classification loss function Domain discriminant loss function Alignment loss function with feature distribution The loss function is obtained by weighted summation. The mathematical expression for the loss function can be found in the following formula.

[0078] (12)

[0079] In the above formula, , and The weighting coefficients are preset and can be arbitrarily set according to actual needs. The weighting coefficients for the classifier, domain discriminator and feature extractor in the network are different to balance the contributions of the three types of loss in the training process.

[0080] The training process involves using one-dimensional single-photon echo data from both the source and target domain training sets as initial model inputs. In each training batch, the data from both domains are processed through forward propagation to obtain predicted labels, domain-specific predictions, and feature vectors. After calculating the classification loss, domain-specific loss, and feature distribution alignment loss, the network parameters are updated via backpropagation. This training process yields optimized model weights for the feature extractor, classifier, and domain discriminator.

[0081] Step S140: After each iteration, the initial model is validated, and the validated initial model is marked as the recognition model and output.

[0082] In one embodiment, step S140: after each iteration, the initial model is validated, including: after each iteration, obtaining the predicted labels output by the initial model processing the training sample set; obtaining the real labels in the training sample set; obtaining the classification accuracy based on the predicted labels and the real labels; when the classification accuracy meets the preset validation conditions, the initial model is deemed to have passed the validation; and / or, after each iteration, obtaining the current iteration round of the initial model; when the current iteration round meets the preset exit conditions, the initial model is deemed to have passed the validation.

[0083] In one implementation, after each iteration, the initial model being trained can be validated to determine whether the current model has reached a preset convergence condition. If it does, training stops and the final model parameters are saved; otherwise, the next iteration continues. Validation can be performed in two ways: first, by obtaining the classification accuracy on the validation set; second, by verifying that the iteration rounds meet the preset exit conditions.

[0084] For the first approach, obtain the predicted labels output from the initial model's training sample set in the current round, and obtain the true labels within the training sample set. Calculate the classification accuracy based on the predicted and true labels, and determine if the classification accuracy meets a preset validation condition. This validation condition could be whether the classification accuracy reaches a preset threshold, or whether the accuracy improvement over several consecutive rounds is less than a certain small value. If either condition is met, the model is considered converged, training is stopped, and the current model parameters are saved; otherwise, continue to the next iteration.

[0085] For the second method, the current iteration count is obtained, and it is determined whether the preset maximum iteration count has been reached. If it has, the model is considered converged, training is stopped, and the final model parameters are saved; otherwise, the next iteration continues. These two verification methods can be used individually or in combination to ensure an optimal balance between performance and efficiency. After training terminates, the initial model is marked as the final, successfully trained recognition model.

[0086] Regarding the recognition model, this application also provides an application method, namely, a single-photon lidar image-free recognition method. For a clear description of this method, please refer to... Figure 4 This includes steps S410 to S440.

[0087] Step S410: Acquire the echo data collected by the single-photon lidar for the target scene.

[0088] In one embodiment, the setup of the single-photon lidar and the acquisition of echo data have been described in detail in step S110 above, and the acquisition method is the same, so it will not be repeated here. Alternatively, acquisition may not be required, and the echo data to be processed can be directly identified and processed.

[0089] Step S420: Input the echo data into the recognition model.

[0090] Step S430: Process the echo data using the recognition model to obtain the recognition result, which includes the pose and type of objects in the target scene.

[0091] In one implementation, the recognition model is trained using the training method described above. For details on the model construction and training process, please refer to the training method provided above; it will not be repeated here. The echo data is directly input into the recognition model for processing to obtain the recognition result, which is used to indicate information such as the pose and type of objects in the target scene.

[0092] To illustrate the advancements of this application, experimental data is provided below; please refer to the relevant documentation for details. Figure 5 , Figure 6 .in Figure 5 This is a schematic diagram of the experimental results of an existing data-driven single-photon radar identification method. The training was conducted on a 200m training dataset and then directly used 300m data. The source domain identification accuracy reached 100%, and the target domain accuracy reached 47.17%. Figure 6 This diagram illustrates the experimental results of the method described in this application under the same conditions, with a source domain of 200m and a target domain of 300m. The source domain recognition accuracy reaches 98.5%, and the target domain accuracy reaches 88.83%. See also... Figure 4 Figure 5 The method in this embodiment can achieve high-quality target recognition even in scenarios without long-distance target labels, and has good robustness.

[0093] Therefore, compared to existing data-driven single-photon radar recognition methods, this application adopts a domain adversarial transfer learning-based single-photon radar recognition technology. It utilizes readily available near-range labeled data as the source domain and combines it with long-range unlabeled data as the target domain for training. This eliminates the need to collect large amounts of labeled data in long-range scenarios, significantly reducing data acquisition costs and labeling workload. Furthermore, the construction of the loss function during training employs multiple loss functions compared to a single loss function, overcoming the limitations of a single loss function's limited scope and weak adaptability, thus achieving comprehensive optimization of model training. This effectively reduces recognition errors caused by distribution differences between the source and target domains, ensuring the model maintains high efficiency and reliable recognition performance even in long-range unlabeled scenarios. Compared to the limitations of traditional multilayer perceptrons and basic convolutional networks in fully mining single-photon echo features, this application improves the model architecture. The feature extractor uses a design that downsamples and compresses features through multilayer encoding paths and upsamples and restores resolution through multilayer decoding paths. This effectively enhances the feature extraction capability for weak echo signals at long distances, improving target recognition accuracy at long distances.

[0094] Figure 7 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 7 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a recognition model training method or a single-photon lidar image-free recognition method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a recognition model training method or a single-photon lidar image-free recognition method. Those skilled in the art will understand that… Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0095] In one embodiment, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any of the foregoing embodiments.

[0096] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0097] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for training a recognition model, characterized in that, Includes the following steps: The system controls a single-photon lidar to emit photon pulses toward a target scene and receives one-dimensional single-photon echoes reflected back from the target scene, wherein the target scene includes at least one object to be identified. Constructing a training sample set based on the one-dimensional single-photon echo; the construction of the training sample set based on the one-dimensional single-photon echo includes: obtaining preset filtering conditions and filtering the one-dimensional single-photon echo; the filtering conditions are distance thresholds for the distance intervals of the one-dimensional single-photon echo corresponding to the identified object; determining one-dimensional single-photon echoes whose interval distance is less than the distance threshold corresponding to the filtering conditions, and setting labels according to the identified object; setting the labeled one-dimensional single-photon echoes as the source domain training set; determining one-dimensional single-photon echoes whose interval distance is greater than or equal to the distance threshold corresponding to the filtering conditions, and setting them as the target domain training set; constructing the training sample set based on the source domain training set and the target domain training set; An initial model is constructed, and the training sample set is input into the initial model for iterative training; the training is used to enable the initial model to learn the deep feature distribution patterns of the target domain training set; The initial model construction, which involves inputting the training sample set into the initial model for iterative training, includes: constructing the initial model, which comprises a feature extractor, a classifier, and a domain discriminator; the feature extractor, using the training sample set as input, extracts feature vectors, which are then used as inputs to the classifier and the domain discriminator; the classifier, using the feature vectors from the source domain training set as input, processes and outputs predicted labels; the domain discriminator, using the feature vectors from the source domain training set and the target domain training set as input, processes and outputs a discrimination result, which indicates whether the feature vector comes from the source domain training set or the target domain training set; inputting the training sample set into the initial model for processing to obtain the feature vectors, predicted labels, and discrimination results; constructing a loss function, using the feature vectors, predicted labels, and discrimination results as inputs to the loss function to calculate training parameters; and updating the initial model using the training parameters. After each iteration, the initial model is validated, and the validated initial model is marked as the recognition model and output.

2. The recognition model training method as described in claim 1, characterized in that, The control of the single-photon lidar to emit photon pulses toward the target scene and to receive one-dimensional single-photon echoes reflected back from the target scene includes: The emission parameters of the single-photon lidar are set, including amplitude and frequency; A transmission pulse signal function is determined based on the transmission parameters, and the transmission pulse signal function is used to control the single-photon lidar to emit photon pulses. Acquire the receiving parameters, and determine the average number of probe photons within one pulse period based on the receiving parameters and the transmitted pulse signal function; Based on the average number of detected photons, calculate the photon arrival probability distribution in each time slot; The arrival probability distribution of all pulse periods is accumulated to generate the one-dimensional single-photon echo.

3. The recognition model training method as described in claim 2, characterized in that, The step of calculating the photon arrival probability distribution in each time slot based on the average number of detected photons includes: The average number of detected photons is modeled using Poisson statistics to obtain the detection probability, which is used to describe the probability of detecting a photon in the nth time slot. Based on the detection probability, the photon arrival probability in each time slot is determined and denoted as the arrival probability. Obtain a preset probability decision threshold, filter all arrival probabilities, and mark the time slots corresponding to arrival probabilities greater than the probability decision threshold as valid time slots; The arrival probabilities corresponding to all valid time slots within a period are summarized to obtain the photon arrival probability distribution for the corresponding period.

4. The recognition model training method as described in claim 1, characterized in that, The construction of the loss function includes: Construct a classification loss function by obtaining the first total number of samples and the true labels in the source domain training set; Obtain the second total number of samples in the target domain training set; construct a domain discrimination loss function based on the second total number of samples and the first total number of samples; Construct a feature distribution alignment loss function based on the first total number of samples and the second total number of samples; The classification loss function, the domain discrimination loss function, and the feature distribution alignment loss function are used for network training.

5. The recognition model training method as described in claim 1, characterized in that, The initial model is validated after each iteration, including: After each iteration, the predicted labels output by the initial model processing the training sample set are obtained; the true labels within the training sample set are obtained; the classification accuracy is obtained based on the predicted labels and the true labels; when the classification accuracy meets preset verification conditions, the initial model is deemed to have passed verification; and / or, After each iteration, the current iteration number of the initial model is obtained; when the current iteration number meets the preset exit conditions, the initial model is deemed to have passed the verification.

6. A single-photon lidar image-free recognition method, characterized in that, Includes the following steps: Acquire echo data collected by single-photon lidar for a target scene; The echo data is input into the recognition model, which is obtained by the training method according to any one of claims 1 to 5; The echo data is processed by the recognition model to obtain the recognition result, which includes the pose and type of objects in the target scene.

7. A computer device, characterized in that, Including processor and memory; The processor is used to execute the computer program stored in the memory to implement the recognition model training method as described in any one of claims 1 to 5, or the single-photon lidar image-free recognition method as described in claim 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the recognition model training method as described in any one of claims 1 to 5, or the single-photon lidar image-free recognition method as described in claim 6.

Images