A method and system for preprocessing and compressing single-photon lidar data

Abstract

The application discloses a kind of single-photon laser radar data preprocessing and compression method and system, belong to single-photon laser radar data processing technical field;The method first acquires target distance image by single-photon laser radar imaging system, provides original data for subsequent processing. Then, for each pixel point of the collected image, a frequency-time histogram is constructed to visually present the distribution of photons in the time dimension. Subsequently, Gaussian fitting and peak filtering operations are performed on the histogram of each pixel point to accurately identify the peak value of valid photon signals, eliminate false signals caused by noise and interference, and improve data quality. Finally, the filtered peak value corresponding to the photon data is input into a multi-attention three-dimensional single-photon compression neural network. Using the network's powerful feature extraction, efficient compression mechanism and flexible output control capability, efficient compression of single-photon laser radar data is achieved, providing strong support for the application of single-photon laser radar technology in multiple fields.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of single-photon lidar data processing, and relates to a single-photon lidar data preprocessing and compression method and system. BACKGROUND

[0002] With the continuous progress of lidar technology, single-photon lidar systems based on single-photon avalanche diode arrays (SPAD) have been widely used in many fields due to their excellent performance. In the field of high-precision three-dimensional imaging, it can accurately capture the three-dimensional shape of objects, providing key data support for industrial manufacturing, cultural heritage protection, etc.; in the automatic driving scene, it can perceive the three-dimensional information of the surrounding environment in real time, helping vehicles make accurate decisions and ensuring driving safety; in the field of topographic mapping, single-photon lidar can efficiently obtain large-area terrain data, providing high-precision basic information for geographic information systems (GIS); in the field of biomedical imaging, it provides a powerful tool for the study of microscopic structures and tissue morphology.

[0003] A significant advantage of such systems is that they can achieve sub-centimeter depth perception even in extremely low light conditions, with extremely high time resolution and detection sensitivity. This allows it to work stably in complex environments and obtain high-quality detection data, providing a solid guarantee for applications in various fields.

[0004] In practical applications, single-photon lidar usually outputs frequency-time histogram data. Specifically, for each pixel point in the field of view, the system will count the arrival time of photons over multiple laser emission periods, and then form a histogram on the time axis. This histogram clearly reflects the detection frequency in different time bins, providing an important basis for analyzing the distance, reflectivity, etc. of the target object.

[0005] However, due to the high timing accuracy of single photons, the time axis is often subdivided into thousands of time bins. For high-resolution fields of view, the original data size of each frame is extremely large, up to millions of data points. In continuous acquisition scenarios, the data volume grows exponentially. This massive amount of data puts a huge pressure on the system, greatly increasing the bandwidth, storage and computing costs. Especially in resource-constrained edge devices or real-time systems, the timeliness and accuracy of data processing are severely challenged, and efficient compression strategies are needed to alleviate data pressure.

[0006] In order to deal with the problem of large amount of single-photon lidar data, traditional compression algorithms such as histogram threshold screening, time domain downsampling, entropy encoding and principal component analysis are applied in data processing. These methods can reduce the data volume to a certain extent, and reduce the pressure of data storage and transmission.

[0007] However, they have many drawbacks in practical application. Especially in processing multi-peak signal, complex background noise or sparse data, these methods often cannot effectively preserve the detailed information of the target echo. Histogram threshold screening may mistakenly filter out some key but small amplitude signal peaks; time domain downsampling will cause the time resolution to be reduced, and part of the important information will be lost; entropy encoding and principal component analysis have limited compression effect when facing complex data structure, and may not be accurate enough for feature extraction of data. These problems make it difficult to recover high-precision target information when the compressed data is reconstructed, which seriously affects the performance and effect of single-photon lidar system in practical application. SUMMARY

[0008] The purpose of the present application is to solve the technical problem that the existing single-photon lidar generates a large amount of redundant data when obtaining high-precision depth information, and it is difficult to perform efficient real-time calculation, and to provide a single-photon lidar data preprocessing and compression method and system.

[0009] In order to achieve the above purpose, the technical scheme is adopted as follows:

[0010] The present application provides a single-photon lidar data preprocessing and compression method in the first aspect, comprising the following steps:

[0011] Collecting target distance images by a single-photon lidar imaging system;

[0012] Constructing a corresponding frequency-time histogram for each pixel point of the collected target distance image;

[0013] Gaussian fitting is performed on the frequency-time histogram of each pixel point, and peak value screening is performed;

[0014] The screened peak value corresponding photon data is input into a multi-attention three-dimensional single-photon compression neural network for data compression.

[0015] Further, the single-photon lidar imaging system adopts a paraxial single-photon lidar system, which includes a laser, a collimating mirror, a single-photon detector, a time-dependent single-photon counting module, a driver and a computer.

[0016] The laser is electrically connected to a time-correlated single-photon counting module, with the laser emission time of the laser as the photon emission time; the laser emitted by the laser is collimated by a collimating lens and then pointed at the target; the reflected light from the target enters the single-photon detector for signal detection;

[0017] The single-photon detector is electrically connected to the time-correlated single-photon counting module, and the time when the single-photon detector receives the laser is taken as the photon reception time.

[0018] A displacement stage is provided at the bottom of the single-photon detector; the driver is electrically connected to the displacement stage, and the driver controls the displacement stage to move the single-photon detector to adjust the target distance and achieve scanning of targets at different distances.

[0019] Furthermore, the step of performing Gaussian fitting on the frequency-time histogram of each pixel and filtering for peak values ​​specifically involves:

[0020] The frequency-time histogram of each pixel is fitted with a Gaussian mixture function to obtain the photon distribution. One fitting function;

[0021] The optimization is achieved using the least squares method. The parameters of the fitting function are obtained. The parameters of a fitting function;

[0022] according to Peak values ​​are selected from the parameters of the fitted function to obtain the target peak value.

[0023] Furthermore, the photon distribution is fitted using a Gaussian mixture function as follows:

[0024]

[0025] in, It is the first The amplitude of the Gaussian peak; It is the first The mean of the Gaussian peaks; It is the first The standard deviation of the Gaussian peak; Represents the location of Mount Gauss; The first image representing the target distance Line 1 The Gaussian fit result of the frequency-time histogram of the pixels in the column; exp() represents the natural exponential function; This represents the total number of Gaussian peaks.

[0026] Furthermore, the optimization of the least squares method The parameters of the fitting function are as follows:

[0027] Updated using least squares method Until the loss function is minimized; the expression for the loss function is:

[0028]

[0029] in, The set of parameters to be optimized and iterated; The actual observed number in the frequency-time histogram One data point; This represents the fitting function based on the Gaussian mixture model distribution; Indicates the first The location of the Gaussian peak.

[0030] Furthermore, the aforementioned according to Peak values ​​are filtered from the parameters of the fitted function to obtain the target peak value. Specifically, when the amplitude of the fitted function is greater than a set threshold, the signal peak value is retained as the target peak value.

[0031] Furthermore, the step of inputting the photon data corresponding to the selected peaks into a multi-attention three-dimensional single-photon compression neural network for data compression specifically involves:

[0032] Photon data within the time range of the target peak are input into a multi-attention three-dimensional single-photon compression neural network for data compression; the time range is:

[0033]

[0034] in, It is the first The mean of the Gaussian peaks; It is the first The standard deviation of the Gaussian peak; Indicates the time range of the target peak.

[0035] Furthermore, the multi-attention three-dimensional single-photon compressed neural network includes a spatiotemporal attention feature extraction module, a deep compression extraction module, and a compression mapping module connected in sequence;

[0036] The spatiotemporal attention feature extraction module includes several sequentially connected 3D convolution and activation units, attention mechanism module, regularization and normalization unit, and 3D max pooling downsampling unit;

[0037] The deep compression extraction module includes a deep compression extraction stacked unit and a downsampling module connected in sequence; the deep compression extraction stacked unit includes a depth convolution unit, a point convolution unit, and a normalization and nonlinear activation unit; the output of the point convolution unit is connected to the output of the normalization and nonlinear activation unit; the deep compression extraction stacked unit is stacked repeatedly several times;

[0038] The compression mapping module sequentially comprises a flattening layer, a multi-layer fully connected transformation unit and a Sigmoid mapping layer; the multi-layer fully connected transformation unit is repeatedly stacked for several times.

[0039] The second aspect of the application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to realize the preprocessing and compression method of the single-photon laser radar data.

[0040] The third aspect of the application provides a preprocessing and compression system of single-photon laser radar data, comprising:

[0041] The data acquisition module collects target distance images through a single-photon laser radar imaging system;

[0042] The histogram construction module constructs a corresponding frequency-time histogram for each pixel point of the collected target distance image;

[0043] The peak value screening module performs Gaussian fitting on the frequency-time histogram of each pixel point and performs peak value screening;

[0044] The data compression module inputs the screened peak value corresponding photon data into a multi-attention three-dimensional single-photon compression neural network for data compression.

[0045] Compared with the prior art, the application has the following beneficial effects:

[0046] The application discloses a preprocessing and compression method of single-photon laser radar data, constructs a frequency-time histogram for each pixel point, converts single-photon data from "overall fuzzy distribution" to "pixel-level accurate timing characteristics", can clearly present the photon response law of each pixel point in different time dimensions, effectively distinguishes the timing difference between target signal photons and background noise photons, Gaussian fitting can accurately model the effective signal trend in the histogram based on the natural distribution characteristics of the photon signal, reduces the interference of random noise on the signal characteristics, peak value screening further focuses on the effective signal peak value after fitting, eliminates the photon data corresponding to low-amplitude noise, significantly improves the signal-to-noise ratio of the photon data, greatly reduces the effective data input into the compression network, reduces data redundancy, and reduces the subsequent compression cost; the multi-attention three-dimensional single-photon compression neural network can automatically identify the key feature area in the single-photon data through the attention mechanism, and allocate more compression resources to the key area, while efficiently compressing the background area with high redundancy, effectively improves the data compression ratio under the premise of ensuring the fidelity of the target data.

[0047] Further, the least square method is used to fit the frequency-time histogram of each pixel point to obtain the peak value of the frequency-time histogram of each pixel point. KThe parameters of the fitting function are optimized, the precise solution of the parameters of the fitting function can be realized based on the error minimization target between the actual observation value of the photon count in the histogram and the predicted value of the fitting function, the interference of random noise on the parameter estimation is effectively reduced, the obtained K The fitting function parameters have higher stability and reliability; based on the optimized K The peak value screening is performed based on the optimized fitting function parameters, the peak value position of each signal component corresponding to the fitting function can be accurately located, and then the peak value corresponding to the target signal is accurately separated from the complex photon distribution, through the analysis of the fitting function parameters, the target signal peak value, background noise peak value and stray light interference peak value can be effectively distinguished, and the final screened target peak value can accurately reflect the distance information of the target, and the influence of invalid noise signals on the subsequent compression link is further reduced. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments, and it should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the premise of the drawings.

[0049] Figure 1 The flow chart of the preprocessing and compression method of the single-photon laser radar data of the present application is shown in the figure.

[0050] Figure 2 The structure diagram of the paraxial single-photon laser radar system of the present application is shown in the figure.

[0051] Figure 3 The photon data graph before coarse screening of the present application is shown in the figure.

[0052] Figure 4 The photon data graph after coarse screening of the present application is shown in the figure.

[0053] Figure 5 The structure diagram of the multi-attention three-dimensional single-photon compression neural network of the present application is shown in the figure.

[0054] Figure 6 The structure diagram of the space-time feature extraction module of the present application is shown in the figure.

[0055] Figure 7 The structure diagram of the deep compression extraction module of the present application is shown in the figure.

[0056] Figure 8 The structure diagram of the compression mapping module of the present application is shown in the figure.

[0057] Figure 9The system block diagram of the pre-processing and compression of the single-photon lidar data.

[0058] 1-time correlation single-photon counting module; 2-laser; 3-collimating mirror; 4-optical lens group; 5-filter; 6-single-photon detector; 7-displacement stage; 8-driver; 9-computer; 901-data acquisition module; 902-histogram construction module; 903-peak screening module; and 904-data compression module. DETAILED DESCRIPTION

[0059] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and indicated in the drawings can be arranged and designed in various different configurations.

[0060] Therefore, the detailed description of the embodiments of the present application provided in the drawings below is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0061] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0062] The present application will be described in further detail below with reference to the drawings:

[0063] Referring to Figure 1 The present application discloses a pre-processing and compression method of single-photon lidar data, comprising the following steps:

[0064] S1, a single-photon lidar imaging system is built:

[0065] Referring to Figure 2 First, a high-efficiency and low-noise paraxial single-photon lidar system is constructed. The paraxial single-photon lidar system comprises a laser 2, a collimating mirror 3, an optical lens group 4, a filter 5, a single-photon detector 6, a time correlation single-photon counting (TCSPC, Time-Correlated Single Photon Counting) module 1 and a displacement stage 7.

[0066] The light emitted by the laser 2 is directed to the target through the collimating mirror 3, and the light reflected by the target is transmitted to the single-photon detector 6 through the optical lens group 4 and the optical filter 5 in turn. The single-photon detector 6 is carried by the displacement table 7 and is controlled to move by the driver 8. The "start" end of the time-correlated single-photon counting module 1 is electrically connected with the laser 2; the "stop" end of the time-correlated single-photon counting module 1 is electrically connected with the single-photon detector 6; the "enable" end of the time-correlated single-photon counting module 1 is electrically connected with the computer 9; and the computer 9 is electrically connected with the driver 8.

[0067] The computer 9 triggers the time-correlated single-photon counting module 1 to start and makes the laser 2 emit a laser pulse. The laser is collimated by the collimating mirror 3 and irradiates the target. The laser reflected by the target is transmitted to the single-photon detector 6 through the optical lens group 4 and the optical filter 5 in turn, is detected by the single-photon detector 6, and the time when the laser is emitted is recorded as "start", as the photon emission time. The time when the single-photon detector 6 receives the laser is recorded as "stop", as the photon receiving time. The time-correlated single-photon counting module 1 records the time difference between "start" and "stop" to calculate the target distance. At the same time, the single-photon detector 6 on the displacement table 7 is controlled to move by the driver 8 to realize scanning of targets at different distances.

[0068] The laser emitted by the laser 2 is irradiated to the surface of the target object through the collimating mirror 3 and is received by the single-photon detector 6 after being reflected by the target. In order to realize the multi-distance imaging capability, a group of displacement tables 7 are introduced to accurately move the position of the single-photon detector 6 under the premise that the position of the front imaging lens is kept fixed, so as to realize the adjustment of the echo focal plane and the expansion of the imaging distance range.

[0069] S2, single-photon lidar data acquisition:

[0070] Referring to Figure 3 The single-photon lidar imaging system is aligned with the measured target, data acquisition is performed, and the target distance image obtained has a size of , wherein, is the number of frames.

[0071] S3, data coarse screening:

[0072] S301, data preprocessing:

[0073] For each pixel point of the collected target distance image , it is expressed as a time sequence in the time dimension , wherein, is the length of the time dimension, respectively represent the time when the photon arrives when the length of the time dimension is ; and and is the coordinate of the pixel point in the target distance image, where represents the row number, represents the column number.

[0074] The size of the acquired target distance image is: , where is the field of view size of the single photon detector, is the length of the time sequence acquired.

[0075] S302, frequency-time histogram construction:

[0076] For the data of each pixel point , first, by setting an appropriate time bin width , the frequency-time histogram of the pixel point is constructed . The histogram represents the number of photon arrivals in different time bins, and the expression is:

[0077]

[0078] where, the axis of the histogram represents time, represents the photon count in the time bin ; and represents the counting function.

[0079] S303, Gaussian mixture function fitting:

[0080] The frequency-time histogram of each pixel point is fitted by a Gaussian mixture function to obtain the main signal peak, and the specific steps are as follows:

[0081] S303-1, Gaussian fitting of the frequency-time histogram of each pixel point is performed by a Gaussian mixture function, and the specific expression of the Gaussian mixture function is:

[0082]

[0083] where, is the amplitude of the th Gaussian peak, representing the peak intensity; is the mean of the th Gaussian peak, representing the position of the Gaussian peak, i.e., the time of photon arrival; is the standard deviation of the th Gaussian peak, representing the width of the peak; is the total number of Gaussian peaks; exp() represents the natural exponential function; Gaussian fitting result of the frequency-time histogram representing the pixel point of the first row and the first column of the target distance image.

[0084] S303-2, using the least square method to optimize the fitting function based on the Gaussian mixture function , wherein represents a set of iterative parameters to be optimized; let:

[0085]

[0086] then:

[0087]

[0088] The partial derivative of each parameter is calculated during the optimization process, and is constantly iteratively updated until the loss function converges. The loss function is expressed as:

[0089]

[0090] wherein, represents the first data point actually observed in the frequency-time histogram; is the fitting function value corresponding to the first data point actually observed in the frequency-time histogram; represents the position of the first Gaussian peak.

[0091] The optimal parameter set obtained by fitting is expressed as:

[0092]

[0093]

[0094] Finally, the parameters of the fitting function are obtained, and each parameter of the fitting function is expressed as ; wherein, argmin represents the value of the independent variable that makes the objective function minimum.

[0095] S304, peak selection and screening:

[0096] According to the fitting result of the Gaussian mixture function, i.e. the parameters of the fitting function, peak screening is performed, and according to the amplitude significant peaks are selected, and only when the amplitude is greater than a set threshold , the peak is retained as a target peak, and the expression is:

[0097]

[0098] This can avoid the reservation of noise peaks and invalid peaks.

[0099] The time range of the target peak is determined by its mean and standard deviation :

[0100]

[0101] The time range contains the main part of the target peak, and the photon data in this time range is used as the input of the subsequent compression process. This time range can be recorded as .

[0102] S305, data output:

[0103] S304 step is obtained by Gaussian mixture model fitting and peak screening. The main peak region of each pixel point is obtained, that is, the time range . The photon data in this time range will be reserved for subsequent data compression processing. Specifically:

[0104] For each pixel point , the photon data in the time range is reserved; the time bin data not in the range is removed, thereby reducing the data size and removing background noise. Finally, the output data will be a compressed time sequence containing the peak region, and the data size is , is expressed as a compressed time sequence, see Figure 4 , and the photon counting cube is finally screened .

[0105] S4, data compression processing:

[0106] The data size after Gaussian fitting coarse screening is , , which represents the number of effective bins extracted in the time dimension. The input data is in the form of a 5-dimensional tensor [Batch, Height, Width, Depth, Channel] sent to the neural network. The default batch size is 1, and the channel number is 1. The input tensor size is: ; Batch represents the sample size; Height represents the image height; Width represents the image width; Depth represents the image depth; Channel represents the image channel number.

[0107] see Figures 5-8The multi-attention three-dimensional single-photon compression neural network comprises a space-time attention feature extraction module, a deep-level compression extraction module and a compression mapping module. The photon data corresponding to the screened peak values are sequentially processed by the space-time attention feature extraction module, the deep-level compression extraction module and the compression mapping module, and finally compressed data is output.

[0108] The core purpose of the space-time attention feature extraction module is to extract discriminative space-time fusion features from the input high-dimensional single-photon three-dimensional tensor, so as to enhance the expression ability of the network to weak signals in single-photon events. The space-time attention feature extraction module receives input data with a dimension of , and outputs intermediate feature data with a dimension of , providing multi-dimensional feature support for subsequent deep compression processing. The space-time attention feature extraction module mainly includes the following structures, which are repeatedly stacked in a stacked form times:

[0109] (1) Three-dimensional convolution and activation unit (Conv3D+ReLU):

[0110] The screened peak value corresponding photon data input is first subjected to a set of three-dimensional convolution operations (Conv3D, 3DConvolutional Layer), and the convolution kernel covers the space and time dimensions, so as to realize the space-time feature perception in the local region. The convolution output is then subjected to a ReLU (Rectified Linear Unit, rectified linear unit) activation function, so as to enhance the nonlinear modeling ability of the network and speed up the convergence speed.

[0111] (2) Attention mechanism module (CBAM, Convolutional Block Attention Module):

[0112] The feature tensor processed by the three-dimensional convolution and activation unit is input into the CBAM module. The CBAM module combines the channel attention and the spatial attention mechanism, respectively models the importance relationship between channels and the saliency area of the spatial position, realizes adaptive enhancement of key area features, and effectively suppresses background noise and redundant signals.

[0113] (3) Regularization and normalization unit (Dropout+LayerNorm):

[0114] Dropout operation is introduced after CBAM output to reduce the risk of overfitting, and then layer normalization (LayerNorm) processing is performed to improve the network stability and training efficiency, which is particularly suitable for processing variable-length time series and weak signal scenarios.

[0115] (4) Three-dimensional maximum pooling downsampling unit (MaxPooling3D):

[0116] The module output is finally realized by three-dimensional maximum pooling to reduce the sampling of the spatial and temporal dimensions, effectively compressing the feature volume and retaining the backbone information, providing more compact feature representation for subsequent deep compression.

[0117] A deep compression extraction module, which adopts a depth separable three-dimensional convolution mechanism, has spatial-time joint modeling capability and high parameter efficiency, and has an input data dimension of and an output data dimension of The deep compression extraction module includes a deep compression extraction stacking unit and a down-sampling module; the deep compression extraction stacking unit is mainly formed by stacking a depth convolution unit, a point convolution unit, and a normalization and nonlinear activation unit in sequence, and specifically:

[0118] (1) Depth convolution unit (Depthwise Conv3D):

[0119] The input tensor is first subjected to a depth convolution operation, that is, a three-dimensional convolution operation is independently performed on each input channel to extract local spatio-temporal features without introducing cross-channel information interaction, significantly reducing the computational burden and parameter size.

[0120] (2) Point convolution unit (Pointwise Conv3D):

[0121] Then, point convolution (i.e., three-dimensional convolution of 1×1×11\times1\times11×1×1) is performed to realize cross-channel feature fusion and improve expression ability. This module allows the channel dimension to be reconstructed while maintaining the size of the input feature map unchanged.

[0122] (3) Normalization and nonlinear activation unit (BatchNorm+ReLU):

[0123] The point convolution output is processed by batch normalization (Batch Normalization) to stabilize the training process, and then a ReLU activation function is introduced to introduce nonlinear mapping capability.

[0124] (4) Multiple stacking and residual connection (repeat×n and jump connection):

[0125] The above three unit modules are repeatedly stacked times to enhance the depth and representation ability of the model; at the same time, a short circuit connection structure is adopted, which directly weights the input of the nonlinear activation unit to the final output of the stacking module to form a residual path, improving the training stability and gradient flow.

[0126] (5) Down-sampling module (MaxPooling3D):

[0127] The output of the residual connection is sent to a three-dimensional max-pooling layer, which is down-sampled in the spatial and temporal dimensions, further compresses the feature size and enhances the modeling capability in the temporal scale.

[0128] The compression mapping module is located at the end of the entire compression neural network, which is used to further map the extracted and compressed multi-dimensional space-time features to a one-dimensional compressed vector to meet the needs of low-bit coding or input reconstruction. The compression mapping module can dynamically adjust the parameter configuration of the fully connected network structure according to the setting of the target compression code length , so as to realize flexible output dimension control and compression ratio adjustment, and balance the compression accuracy and reconstruction fidelity. Specifically, the structure of the compression mapping module is as follows:

[0129] (1) Flatten Layer:

[0130] Firstly, the input five-dimensional tensor is unfolded into a one-dimensional vector, which breaks through the structural barriers among space, time and channel, and prepares for the subsequent fully connected mapping.

[0131] (2) Multi-layer fully connected transformation unit (Fully Connected Layer + ReLU):

[0132] The unfolded high-dimensional vector is input into multiple fully connected layers (Fully Connected, FC) in turn, each of which is followed by a ReLU activation function to introduce non-linear expression capability. This part can be repeatedly stacked times to further compress and refine the high-dimensional features.

[0133] (3) Sigmoid mapping layer:

[0134] The last layer is a Sigmoid activation function, which is used to compress the output range to interval, which is suitable for the normalized output form required by the binary coding, photon intensity probability modeling or downstream objective function.

[0135] (4) Dynamic dimension regulation mechanism:

[0136] The final output dimension of the multi-layer fully connected transformation layer can be dynamically set according to the compression target parameters , so as to realize the output compressed vector size . Through this mechanism, the system can flexibly adapt to different transmission bandwidth, storage resources or accuracy requirements.

[0137] In this step, the input end of the space-time attention feature extraction module receives the filtered photon count cube The deep compression extraction module is used to filter the photon counting cube through three-dimensional convolution and attention mechanisms. Compressed into a photon counting matrix and the photon counting matrix The input is sent to the compression mapping module; the compression mapping module is used to process the photon counting matrix through a flattening layer and a fully connected layer. Perform compression mapping to compressed data This refers to compressed single-photon data. The compressed single-photon data... Compared with the filtered single-photon data The following quantitative relationships exist between them:

[0138]

[0139] To enable flexible adjustment of the compression ratio, the compression mapping module includes: A fully connected layer, A positive integer used to boost the compression target parameter in the output dimension. Given the feature mapping capability, to ensure the integrity of the representation of input single-photon data under high compression ratio, the compression ratio is... for: .

[0140] See Figure 9 One embodiment of the present invention provides a preprocessing and compression system for single-photon lidar data, comprising:

[0141] Data acquisition module 901 acquires target distance images through a single-photon lidar imaging system;

[0142] Histogram construction module 902 constructs a frequency-time histogram for each pixel of the acquired target distance image;

[0143] Peak filtering module 903 performs Gaussian fitting on the frequency-time histogram of each pixel and then performs peak filtering.

[0144] The data compression module 904 inputs the photon data corresponding to the selected peaks into a multi-attention three-dimensional single-photon compression neural network for data compression.

[0145] One embodiment of the present application provides a storage medium, specifically a computer readable storage medium, which is a memory device in a terminal device, used for storing programs and data. It can be understood that the computer readable storage medium herein can include a built-in storage medium in the terminal device, and of course can include an expansion storage medium supported by the terminal device, and can be any tangible medium containing or storing programs, which can be used by or in combination with an instruction execution system, device or apparatus. The computer readable storage medium provides a storage space, which stores an operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more computer programs (including program codes). It should be noted that more specific examples (non-exhaustive list) of the computer readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical fiber, a portable compact disk read-only memory, an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0146] The computer readable storage medium also includes a data signal carried in baseband or propagated as a carrier wave in a propagated data signal, in which a readable program code is borne. Such a propagated data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The readable storage medium can also be any readable medium other than the readable storage medium, which can send, propagate or transmit programs for use by or in combination with an instruction execution system, device or apparatus. The program code contained on the readable storage medium can be transmitted by any suitable medium, including but not limited to wireless, wired, optical cable, etc., or any suitable combination of the above.

[0147] The program code for performing the operations of the present application can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, etc., and a conventional procedural programming language such as the "C" programming language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any kind of network, including a local area network or a wide area network, or can be connected to an external computing device (for example, through the Internet by connecting to an Internet service provider).

[0148] The one or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the preprocessing and compression method of the single-photon lidar data in the above embodiments.

[0149] The above only is the preferred embodiment of the present application, and is not used to limit the present application, for the person skilled in the art, the present application can have various changes and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for preprocessing and compressing single-photon lidar data, characterized in that, Includes the following steps: Target distance image acquisition is performed using a single-photon lidar imaging system; For each pixel in the acquired target distance image, construct the corresponding frequency-time histogram; Gaussian fitting was performed on the frequency-time histogram of each pixel, and peak values ​​were filtered out. The photon data corresponding to the selected peaks are input into a multi-attention 3D single-photon compression neural network for data compression. Specifically, the process of performing Gaussian fitting on the frequency-time histogram of each pixel and filtering for peak values ​​involves: The frequency-time histogram of each pixel is fitted with a Gaussian mixture function to obtain the photon distribution. A fitting function; specifically, the photon distribution is fitted using a Gaussian mixture function: in, It is the first The amplitude of the Gaussian peak; It is the first The mean of the Gaussian peaks; It is the first The standard deviation of the Gaussian peak; Represents the location of Mount Gauss; The first image representing the target distance Line 1 The Gaussian fit result of the frequency-time histogram of the pixels in the column; exp() represents the natural exponential function; This represents the total number of Gaussian peaks; The optimization is achieved using the least squares method. The parameters of the fitting function are obtained. The parameters of the fitting function are: Updated using least squares method Until the loss function is minimized; the expression for the loss function is: in, The set of parameters to be optimized and iterated; The actual observed number in the frequency-time histogram One data point; This represents the fitting function based on the Gaussian mixture model distribution; Indicates the first The location of the Gaussian peak; according to Peak values ​​are selected from the parameters of the fitted function to obtain the target peak value; The step of inputting the photon data corresponding to the selected peaks into a multi-attention three-dimensional single-photon compression neural network for data compression is as follows: Photon data within the time range of the target peak are input into a multi-attention three-dimensional single-photon compression neural network for data compression; the time range is: in, It is the first The mean of the Gaussian peaks; It is the first The standard deviation of the Gaussian peak; Indicates the time range of the target peak.

2. The method for preprocessing and compressing single-photon lidar data according to claim 1, characterized in that, The single-photon lidar imaging system adopts a side-axis single-photon lidar system, which includes a laser, a collimating lens, a single-photon detector, a time-correlated single-photon counting module, and a driver. The laser is electrically connected to a time-correlated single-photon counting module, with the laser emission time of the laser as the photon emission time; the laser emitted by the laser is collimated by a collimating lens and then pointed at the target; the reflected light from the target enters the single-photon detector for signal detection; The single-photon detector is electrically connected to the time-correlated single-photon counting module, and the time when the single-photon detector receives the laser is taken as the photon reception time. A displacement stage is provided at the bottom of the single-photon detector; the driver is electrically connected to the displacement stage, and the driver controls the displacement stage to move the single-photon detector to adjust the target distance and achieve scanning of targets at different distances.

3. The method for preprocessing and compressing single-photon lidar data according to claim 1, characterized in that, According to Peak values ​​are filtered from the parameters of the fitted function to obtain the target peak value. Specifically, when the amplitude of the fitted function is greater than a set threshold, the signal peak value is retained as the target peak value.

4. The method for preprocessing and compressing single-photon lidar data according to claim 1, characterized in that, The multi-attention three-dimensional single-photon compressed neural network includes a spatiotemporal attention feature extraction module, a deep compression extraction module, and a compression mapping module connected in sequence. The spatiotemporal attention feature extraction module includes several sequentially connected 3D convolution and activation units, attention mechanism module, regularization and normalization unit, and 3D max pooling downsampling unit; The deep compression extraction module includes a deep compression extraction stacked unit and a downsampling module connected in sequence; the deep compression extraction stacked unit includes a depth convolution unit, a point convolution unit, and a normalization and nonlinear activation unit; the output of the point convolution unit is connected to the output of the normalization and nonlinear activation unit; the deep compression extraction stacked unit is stacked repeatedly several times; The compression mapping module includes, in sequence, a flattening layer, a multi-layer fully connected transformation unit, and a Sigmoid mapping layer; the multi-layer fully connected transformation unit is repeatedly stacked several times.

5. 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 preprocessing and compression method for single-photon lidar data as described in any one of claims 1-4.

6. A preprocessing and compression system for single-photon lidar data, based on the preprocessing and compression method for single-photon lidar data according to any one of claims 1-4, characterized in that, include: The data acquisition module acquires target distance images through a single-photon lidar imaging system; The histogram construction module constructs a frequency-time histogram for each pixel in the acquired target distance image; The peak filtering module performs Gaussian fitting on the frequency-time histogram of each pixel and then filters the peak values. The data compression module inputs the photon data corresponding to the selected peaks into a multi-attention 3D single-photon compression neural network for data compression.

Images