Deep learning based gm-apd lidar data augmentation method and system

Abstract

The application discloses a Gm-APD laser radar data enhancement method and system based on deep learning, and relates to the field of data enhancement of laser imaging radars.The application solves the problem that the existing data enhancement method can only be used for single type of data enhancement and cannot realize paired data enhancement.The method combines a reinforcement generative adversarial network and a Pix2Pix network, and comprises the following steps: adding a CBAM attention module to a traditional generative adversarial network, training the added generative adversarial network, and obtaining a reinforcement generative adversarial network; generating an enhanced intensity image according to the reinforcement generative adversarial network; training a Pix2Pix network using paired intensity images and distance images to generate a Pix2Pix network optimization model; and inputting the enhanced intensity image into the Pix2Pix network optimization model to obtain an enhanced distance image corresponding to the enhanced intensity image.The application is suitable for the field of dynamic vehicle detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data enhancement of laser imaging radar, and particularly to a Gm-APD laser radar data enhancement method based on deep learning. BACKGROUND

[0002] With the rapid development of the field of laser radar, as an active imaging technology, laser radar has become a research hotspot in the field of target detection. Compared with other passive detection methods such as near-infrared imaging and visible light imaging, laser imaging radar is an active detection method, which is less affected by external environmental factors such as weather, climate and background light, and can detect more rich target information, including surface reflectivity, speed, surface features, etc. It provides more target features and information support for detection, tracking and identification applications, and is widely used in target identification, ground remote sensing observation, city three-dimensional modeling, robot and unmanned vehicle obstacle avoidance, navigation and other fields. Among them, Geiger-mode Avalanche Photo Diode (Gm-APD) has single-photon detection capability, and Gm-APD laser radar based on this principle has the characteristics of high sensitivity, fast response speed and high integration of detectors. This kind of laser radar has good application prospect in the above fields.

[0003] In recent years, deep learning technology has become a research hotspot and has developed rapidly. At present, most of the fields where laser radar is widely applied are combined with the data collected by laser radar equipment and deep learning technology, and a large number of samples are needed for network training through deep learning method. When the sample data is less, the training model effect is not good. However, in view of the high cost of obtaining samples by existing laser radar and the problem of less diversity of obtaining samples, it is of great significance to propose an effective data enhancement method for the application of deep learning technology.

[0004] The Gm-APD laser radar data not only contains range image information of target distance, but also includes intensity image information with target reflection characteristics. Both types of data information are equally important, and the two types of data are one-to-one corresponding paired data. The existing data enhancement method can only enhance single type of data, and cannot realize paired data enhancement. SUMMARY

[0005] The present application solves the problem that the existing data enhancement method can only enhance single type of data and cannot realize paired data enhancement.

[0006] The present application provides a Gm-APD laser radar data enhancement method based on deep learning, which combines reinforcement generative adversarial network and Pix2Pix network, comprising:

[0007] adding the CBAM attention module to the traditional generative adversarial network, training the added generative adversarial network to obtain a reinforced generative adversarial network;

[0008] generating an enhanced intensity image according to the reinforced generative adversarial network;

[0009] training a Pix2Pix network using the paired intensity image and distance image to generate a Pix2Pix network optimization model;

[0010] inputting the enhanced intensity image into the Pix2Pix network optimization model to obtain an enhanced distance image corresponding to the enhanced intensity image.

[0011] Further, a preferred mode is also provided, and the training of the generative adversarial network after adding the attention module includes:

[0012] training the generative adversarial network after adding the attention module, wherein the generative adversarial network after adding the attention module includes a generative model G and a discriminative model D;

[0013] training the generative model G and the discriminative model D, including:

[0014]

[0015] wherein E x～P(x) represents the mathematical expectation of x from the real sample, E z～P(z) represents the mathematical expectation of the simulation sample generated by the generator through the noise z, D(x) represents the probability of x being classified as a real sample, if the sample comes from the real training data x, D(x) considers x as true, and outputs a score close to 1, otherwise considers x as false, and outputs a score close to 0. 1-D[G(z)] is the probability of judging the simulation sample G(z) as false.

[0016] Further, a preferred mode is also provided, and the training of the Pix2Pix network using the paired intensity image and distance image to generate the Pix2Pix network optimization model includes:

[0017] collecting vehicle data of the Gm-APD laser radar for providing the paired intensity image and distance image;

[0018] inputting the vehicle data into the Pix2Pix network to obtain the Pix2Pix network optimization model.

[0019] Further, a preferred mode is also provided, and the obtaining of the enhanced distance image corresponding to the enhanced intensity image specifically includes:

[0020] L L1 (G)=E x,y,z[||y-G(x,z)||1],

[0021] wherein L L1 (G) is a loss function of the generator, y is a real image, G(x,z) is a generated image, E x,y,z is to calculate the difference between the real image and the generated image.

[0022] Based on the same inventive concept, the application also provides a deep learning-based Gm-APD laser radar data enhancement system, which combines a reinforced generative adversarial network and a Pix2Pix network, and comprises:

[0023] A reinforced generative adversarial network acquisition unit is configured to add a CBAM attention module to a conventional generative adversarial network, train the generative adversarial network after adding the attention module, and obtain a reinforced generative adversarial network.

[0024] An enhanced intensity image acquisition unit is configured to generate an enhanced intensity image based on the reinforced generative adversarial network.

[0025] A Pix2Pix network optimization model acquisition unit is configured to train a Pix2Pix network using a pair of intensity images and distance images to generate a Pix2Pix network optimization model.

[0026] An enhanced distance image acquisition unit is configured to input the enhanced intensity image into the Pix2Pix network optimization model to obtain an enhanced distance image corresponding to the enhanced intensity image.

[0027] Further, a preferred mode is provided, wherein the reinforced generative adversarial network acquisition unit comprises:

[0028] The generative adversarial network after adding the attention module comprises a generative model G and a discriminative model D.

[0029] The training of the generative model G and the discriminative model D comprises:

[0030]

[0031] wherein E x～P(x) represents the mathematical expectation of x from a real sample, E z～P(z) represents the mathematical expectation of a simulated sample generated by the generator through noise z, D(x) represents the probability of x being classified as a real sample, if the sample comes from a real training data x, D(x) considers x as true, outputs a score close to 1, otherwise considers x as false, and outputs a score close to 0. 1-D[G(z)] is the probability of judging the simulated sample G(z) as false.

[0032] Further, the application further provides a preferred mode of the Pix2Pix network optimization model acquisition unit, which comprises:

[0033] Vehicle data of the Gm-APD laser radar is collected to provide paired intensity images and distance images;

[0034] The vehicle data is input into the Pix2Pix network to acquire a Pix2Pix network optimization model.

[0035] Further, the application further provides a preferred mode of the enhanced distance image data acquisition unit, which specifically comprises:

[0036] L L1 (G)=E x,y,z [||y-G(x,z)||1],

[0037] Wherein, L L1 (G) is a loss function of the generator, y is a real image, G(x,z) is a generated image, and E x,y,z is used to calculate the difference between the real image and the generated image.

[0038] Based on the same inventive concept, the application further provides a computer readable storage medium for storing a computer program, wherein the computer program executes the deep learning-based Gm-APD laser radar data enhancement method according to any one of the above.

[0039] Based on the same inventive concept, the application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the deep learning-based Gm-APD laser radar data enhancement method according to any one of the above.

[0040] The application has the advantages that:

[0041] The application solves the problem that the existing data enhancement method can only be used for single type of data enhancement and cannot realize paired data enhancement.

[0042] The Gm-APD lidar data enhancement method based on deep learning provided in the application can improve the feature extraction capability of the network by adding an attention mechanism module in view of the problem of insufficient feature extraction capability and unstable training of a traditional generative adversarial network (GAN).

[0043] The CBAM attention mechanism module is added to the traditional GAN network, and the attention feature map information in the channel and space dimensions is effectively utilized, the target information is enhanced while the noise information is suppressed, the performance of the network is improved, and a reinforced generative adversarial network is obtained.

[0044] The Gm-APD lidar data enhancement method based on deep learning provided in the application can simultaneously enhance the range image and intensity image of the target, solve the problem of being unable to enhance paired data, improve the diversity of the target sample size, save acquisition resources, and provide diversified data support for the training of deep learning for lidar target detection.

[0045] The application is suitable for the field of dynamic vehicle detection and can also be applied in the fields of lidar image classification and semantic segmentation. Data enhancement can improve the sample size and diversity of the data set, and thus improve the target detection accuracy, image classification accuracy and semantic segmentation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 Flow chart of the deep learning-based Gm-APD lidar data enhancement method according to embodiment one;

[0047] Figure 2 Structure diagram of the CBAM attention module according to embodiment two;

[0048] Figure 3 Structure diagram of the improved GAN network according to embodiment two;

[0049] Figure 4 Structure diagram of the Pix2Pix network according to embodiment four;

[0050] Figure 5 Original image and enhanced image according to embodiment eleven, wherein, Figure 5 (a) is an original intensity image and a distance image; Figure 5 (b) is an enhanced intensity image and a distance image;

[0051] Figure 6 Target scene and enhancement effect diagram according to embodiment eleven, wherein, Figure 6 (a) is a target scene diagram, Figure 6 (b) is an enhanced intensity image, Figure 6 (c) is an enhanced distance image. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments of the present application.

[0053] Embodiment one, see Figure 1 This embodiment describes a deep learning-based Gm-APD lidar data enhancement method. The method combines a reinforced generative adversarial network and a Pix2Pix network, and includes:

[0054] Add a CBAM attention module to a traditional generative adversarial network, train the generative adversarial network with the added attention module, and obtain a reinforced generative adversarial network;

[0055] Generate an enhanced intensity image according to the reinforced generative adversarial network;

[0056] Train a Pix2Pix network with paired intensity images and distance images to generate a Pix2Pix network optimization model;

[0057] The enhanced intensity image is input into the Pix2Pix network optimization model to obtain an enhanced distance image corresponding to the enhanced intensity image.

[0058] The deep learning-based Gm-APD lidar data enhancement method of the embodiment can make the generated vehicle edges clearer and the distribution of the vehicles on the image closer to the real scene for the long-distance small target scene of the lidar by adding an attention mechanism module to the traditional generative adversarial network (GAN) and considering the positional relationship between the target and the background.

[0059] The deep learning-based Gm-APD lidar data enhancement method combines the enhanced generative adversarial network GAN with the Pix2Pix network, can simultaneously enhance the distance image and the intensity image of the target, solves the problem of being unable to enhance paired data, improves the diversity of the target sample size, saves acquisition resources, and provides diversified data support for the training of deep learning in the field of lidar target detection and other fields.

[0060] The embodiment can improve the feature extraction capability of the network by adding an attention mechanism module to solve the problem of insufficient feature extraction capability and unstable training of the traditional generative adversarial network (GAN). The existing technology uses a Squeeze and excitation (SE) channel attention module to selectively emphasize information features and suppress useless features by learning to use global information to extract effective features. However, this attention mechanism does not consider the spatial features of the data and cannot extract features using spatial information. Therefore, the embodiment adds a Convolutional Block Attention Module (CBAM) attention module to the traditional GAN network, effectively uses the attention feature map information in the channel and spatial dimensions, enhances the target information while suppressing noise information, improves the performance of the network, and thus obtains an enhanced generative adversarial network. The enhanced generative adversarial network is trained using the intensity image of the Gm-APD lidar, and an enhanced intensity image data is generated.

[0061] To solve the problem that the existing data enhancement method cannot enhance paired lidar data, the enhanced generative adversarial network is combined with the Pix2Pix network, the Pix2Pix network is trained using paired intensity image data and distance image data to generate a Pix2Pix network optimization model, and then the generated enhanced intensity image data is input into the Pix2Pix network optimization model to finally obtain an enhanced distance image corresponding to the enhanced intensity image.

[0062] Embodiment two, see Figure 2 and Figure 3 This embodiment is described. This embodiment is a further limitation of the deep learning-based Gm-APD lidar data enhancement method described in embodiment one, the generated adversarial network after adding the attention module, comprising:

[0063] Train the generated adversarial network with the added attention module, which includes a generation model G and a discrimination model D;

[0064] Training the generation model G and the discrimination model D, comprising:

[0065]

[0066] Wherein, E x～P(x) represents the mathematical expectation of x from the real sample, E z～P(z) represents the mathematical expectation of the simulation sample generated by the generator, D(x) represents the probability of x being classified as a real sample, if the sample comes from the real training data x, D(x) then x is true, output close to 1 score, otherwise x is false, output close to 0 score. 1-D[G(z)] is the probability of judging the simulation sample G(z) as false.

[0067] Specifically, the details of the traditional generative adversarial network GAN generating high-resolution images only use the local spatial information of low-resolution feature maps. By adding attention mechanism in the feature extraction part of the network. In 2018, CBAM (Convolutional Block Attention Module Network) attention module was proposed, including spatial attention module SAM (Spatial Attention Module) and channel attention module CAM (Channel Attention Module). They respectively summarize the attention information from the aspects of space and channel, and to a certain extent, the information is integrated, so as to obtain more comprehensive and reliable attention information. The structure is as shown in Figure 2 .

[0068] Adding CBAM attention module to the traditional GAN network can make the whole generative adversarial network consider the relationship between various structures of the target, and consider the position relationship between the target and the background. The intensity image of the laser radar can generate target outline clearer, and the data distribution is closer to the real data distribution. The improved network structure of adding CBAM module to GAN network is as shown in Figure 3 .

[0069] As Figure 3As shown, the network inputs a 100-dimensional Gaussian distribution noise vector, transforms the dimension, then passes through three convolutions, and then adds a CBAM attention module, which introduces an attention mechanism that can enable the network to learn the spatial information and structural information of the target. The network continues to pass through the convolution and CBAM module to extract target features. In order to associate the context information of the target, the results of the conv4 convolution are combined with the conv2 features of the shallow layer, effectively fusing deep features and shallow features, enabling the network to better extract target features and generate more realistic and diverse data.

[0070] The traditional generative adversarial network GAN includes a generative model G and a discriminative model D. The input of the entire model is a high-dimensional random noise (such as Gaussian noise), the generative model G is used to generate simulation samples, and the discriminative model D is used to judge the true class of the sample (only to judge true or false). For the generative model, the greater the loss function of the discriminator, the more difficult it is for the discriminator to distinguish between real samples and simulation samples, indicating that the simulation samples generated by it are closer to real samples. For the discriminative model, the smaller the loss function of the discriminator, the better the discriminative ability of the discriminator for real samples and simulation samples, indicating that the effect of the discriminator is better.

[0071] The training of the generative adversarial network with an attention module includes a generative model G and a discriminative model D, and the training of D and G can be represented as:

[0072]

[0073] wherein E x～P(x) represents the mathematical expectation that x comes from a real sample, E z～P(z) represents the mathematical expectation of the simulation sample generated by the generator from the noise z, D(x) represents the probability that x is classified as a real sample, if the sample comes from a real training data x, D(x) considers x as true, and outputs a score close to 1, otherwise considers x as false, and outputs a score close to 0. 1-D[G(z)] is the probability of judging the simulation sample G(z) as false.

[0074] The network D is trained to maximize the probability of labeling the training sample (maximize log D(x)) and log(1-G(z))), and the network G is trained to minimize log(1-G(z))), that is, to maximize the loss of D. During the training process, one side is fixed, and the parameters of the other network are updated, and the two are iterated alternately, so that the error of the other side is maximized. Finally, G can estimate the distribution of sample data, that is, the generated sample is more realistic.

[0075] Embodiment three, the embodiment is further limited to the Gm-APD lidar data enhancement method based on deep learning of embodiment one, the Pix2Pix network is trained by using the pair of intensity image and distance image to generate the Pix2Pix network optimization model, comprising:

[0076] Collecting vehicle data of Gm-APD lidar for providing the pair of intensity image and distance image;

[0077] The vehicle data is input into the Pix2Pix network, and the Pix2Pix network optimization model is obtained.

[0078] In actual training, specifically:

[0079] Collecting 1552 pairs of vehicle data of Gm-APD lidar;

[0080] The intensity image in the vehicle data is input into the enhanced generative adversarial network, and the enhanced generative adversarial network model is obtained;

[0081] According to the enhanced generative adversarial network model, enhanced intensity image is generated;

[0082] The vehicle data is used to train the Pix2Pix network, and the enhanced Pix2Pix network optimization model is obtained;

[0083] According to the enhanced Pix2Pix network optimization model, the enhanced intensity image is processed to generate the corresponding enhanced distance image.

[0084] Embodiment four, see Figure 4 This embodiment is described. The embodiment is further limited to the enhanced- of embodiment one, and the enhanced distance image corresponding to the enhanced intensity image is obtained, specifically:

[0085] L L1 (G)=E x,y,z [||y-G(x,z)||1],

[0086] Wherein, L L1 (G) is the loss function of the generator, y is the real image, G(x,z) is the generated image, E x,y,z is the difference before calculating the real image and the generated image.

[0087] Specifically, since the Gm-APD lidar image not only contains intensity image, but also contains distance image, only the improved GAN network can enhance single image, therefore, the Pix2Pix network is used to enhance the intensity image after GAN enhancement and distance image, and the Pix2Pix network structure mainly includes two parts, generator and discriminator. The Pix2Pix network structure is as follows Figure 4The generator G uses a Unet structure, and the input profile x is encoded and decoded into a real picture. The discriminator D uses a conditional discriminator PatchGAN. Compared with the traditional GAN discriminator, PatchGAN models the image as a Markov random field, has fewer parameters, runs faster, and can be applied to images of any size. The role of the discriminator D is to judge the generated picture G(x) as false and the real picture as true under the condition of the profile x.

[0088] The loss function of the Pix2Pix network structure is:

[0089]

[0090] wherein, L L1 (G) is the loss function of the generator, λ is a balance constant, L cGAN (G,D) is the loss function of the conditional GAN, and is expressed as:

[0091] L cGAN (G,D)=E x,y [log D(x,y)]+E x,z [log(1-D(x,G(x,z))]

[0092] A regularization term is added to the loss function to improve the quality of the generated image. The difference is that the Pix2Pix network structure uses L1 regularization and does not use L2 regularization. L2 is based on Gaussian prior, and L1 is based on Laplace prior. L1 regularization can reduce image blur and preserve more image edges compared with L2 regularization, as shown in the following formula:

[0093] L L1 (G)=E x,y,z [||y-G(x,z)||1]

[0094] Using L1 regularization helps to make the generated image clearer.

[0095] Embodiment five, the deep learning-based Gm-APD laser radar data enhancement system of the embodiment combines reinforcement generative adversarial network and Pix2Pix network, and the enhancement system comprises:

[0096] The reinforcement generative adversarial network acquisition unit is used to add a CBAM attention module to a traditional generative adversarial network, train the generative adversarial network after adding the attention module, and obtain a reinforcement generative adversarial network.

[0097] The enhanced intensity image acquisition unit is used to generate an enhanced intensity image according to the reinforcement generative adversarial network.

[0098] The Pix2Pix network optimization model acquisition unit is configured to train a Pix2Pix network using the preset intensity image and the preset range image to generate a Pix2Pix network optimization model.

[0099] The enhanced range image acquisition unit is configured to input the enhanced intensity image into the Pix2Pix network optimization model to obtain an enhanced range image corresponding to the enhanced intensity image.

[0100] In an embodiment six, the embodiment is a further limitation of the deep learning-based Gm-APD lidar data enhancement system of the embodiment five, and the reinforcement generative adversarial network acquisition unit comprises:

[0101] The training of the generative adversarial network with an attention module comprises a generative model G and a discriminative model D.

[0102] The training of the generative model G and the discriminative model D comprises:

[0103]

[0104] wherein E x～P(x) represents the mathematical expectation of x from a real sample, E z～P(z) represents the mathematical expectation of a simulated sample generated by the generator through noise z, D(x) represents the probability of x being classified as a real sample, if the sample comes from a real training data x, D(x) considers x as true, and outputs a score close to 1, otherwise considers x as false, and outputs a score close to 0. 1-D[G(z)] is the probability of judging the simulated sample G(z) as false.

[0105] In an embodiment seven, the embodiment is a further limitation of the deep learning-based Gm-APD lidar data enhancement system of the embodiment five, and the Pix2Pix network optimization model acquisition unit comprises:

[0106] The vehicle data of the Gm-APD lidar is collected to provide pairs of preset intensity images and preset range images.

[0107] The vehicle data is input into the Pix2Pix network to obtain a Pix2Pix network optimization model.

[0108] In an embodiment eight, the embodiment is a further limitation of the deep learning-based Gm-APD lidar data enhancement system of the embodiment five, and the range image data acquisition unit specifically comprises:

[0109] L L1 (G)=E x,y,z [||y-G(x,z)||1],

[0110] wherein, L L1 (G) is a loss function of the generator, y is a real image, G(x,z) is a generated image, E x,y,z is to calculate the difference between the real image and the generated image.

[0111] Embodiment nine, a computer readable storage medium according to the embodiment, the computer readable storage medium is used to store a computer program, the computer program executes the Gm-APD laser radar data enhancement method based on deep learning according to any one of the embodiments one to four.

[0112] Embodiment ten, a computer device, comprising a memory and a processor, the memory has stored therein a computer program, when the processor runs the computer program stored in the memory, the processor executes the Gm-APD laser radar data enhancement method based on deep learning according to any one of the embodiments one to four.

[0113] Embodiment eleven, see Figure 5 and Figure 6 This embodiment is described. The embodiment provides a specific embodiment for the Gm-APD laser radar data enhancement method based on deep learning according to the embodiment one, and is also used to explain the embodiments two to four, in particular:

[0114] The experimental environment of the embodiment is: the GPU uses GeForce RTX 2080Ti, the CPU is Intel Core i7-8700CPU@3.20GHz, the hardware acceleration uses CUDA10.0, and the deep learning framework is Pytorch.

[0115] The embodiment fixes the position of the Gm-APD laser radar system of 64x64 pixels, collects the dynamic vehicle data of about 1km on the road at night, collects the vehicle data at night, and the training data is 1552 pairs of intensity images and distance images.

[0116] The intensity image in the 1552 pairs of training data in the night dynamic vehicle data is input into the reinforcement generative adversarial network, and the reinforcement generative adversarial network model is obtained;

[0117] According to the reinforcement generative adversarial network model, 1000 new intensity images are generated;

[0118] The 1552 pairs of training data are used to train the Pix2Pix network, and the Pix2Pix network optimization model is obtained;

[0119] According to the Pix2Pix network optimization model, 1000 intensity images are processed to generate corresponding distance images, and finally 1000 pairs of enhanced distance images and enhanced intensity images are generated, which enhances the diversity of the original target. As shown in Figure 5 , it can be seen that the intensity image and the distance image are enhanced.

[0120] For night dynamic vehicle detection, the target scene and the enhancement effect are as shown in Figure 6 .

[0121] The embodiment combines the enhanced generative adversarial network GAN and the Pix2Pix network to simultaneously enhance the distance image and the intensity image of the Gm-APD laser radar, and increases the diversity of the original data.

[0122] The technical solutions of the present application provided in the above description in combination with the drawings are further described in detail, which is to highlight the advantages and benefits, and is not used as a limitation on the present application. Any modification, combination, improvement and equivalent replacement of the present application based on the principle of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based Gm-APD lidar data enhancement method, characterized in that, The method adopts a combination of a reinforced generative adversarial network and a Pix2Pix network, and comprises the following steps: adding a CBAM attention module to a traditional generative adversarial network, training the generative adversarial network after adding the attention module, and obtaining a reinforced generative adversarial network; generating an enhanced intensity image according to the reinforced generative adversarial network; training a Pix2Pix network using a pair of intensity images and distance images to generate a Pix2Pix network optimization model, comprising the following steps: collecting vehicle data of a Gm-APD laser radar to provide the pair of intensity images and distance images; inputting the vehicle data into the Pix2Pix network to obtain the Pix2Pix network optimization model; inputting the enhanced intensity image into the Pix2Pix network optimization model to obtain an enhanced distance image corresponding to the enhanced intensity image; The Pix2Pix network structure mainly comprises two parts, a generator and a discriminator. The generator G uses a Unet structure, and the input profile x is encoded and then decoded into a real picture.

2. The deep learning-based Gm-APD lidar data augmentation method of claim 1, wherein, The discriminator D uses a conditional discriminator PatchGAN. The training of the generative adversarial network after adding the attention module comprises the following steps: training the generative adversarial network after adding the attention module, wherein the generative adversarial network after adding the attention module comprises a generative model G and a discriminative model D; ， wherein denotes the mathematical expectation from real samples, denotes noise z the mathematical expectation from simulated samples generated by the generator, D ( x ) denotes x the probability that a sample is classified as real if it comes from real training data x , D ( x ) is considered true if is close to 1, otherwise x is considered false and outputs a score close to 0, 1- D [ G ( z ) is the probability that a simulated sample G ( z ) is judged false.

3. The deep learning-based Gm-APD lidar data augmentation method of claim 1, wherein, training the generative model G and the discriminative model D, comprising the following steps: ， wherein, is a loss function for the generator, is a real image, is a generated image, is a difference between the real image and the generated image.

4. A deep learning based Gm-APD lidar data augmentation system, characterized by, The enhanced distance image acquisition unit specifically comprises the following steps: The system adopts a combination of a reinforced generative adversarial network and a Pix2Pix network, and the enhancement system comprises the following units: a reinforced generative adversarial network acquisition unit, configured to add a CBAM attention module to a traditional generative adversarial network, train the generative adversarial network after adding the attention module, and obtain a reinforced generative adversarial network; an enhanced intensity image acquisition unit, configured to generate an enhanced intensity image according to the reinforced generative adversarial network; a Pix2Pix network optimization model acquisition unit, configured to train a Pix2Pix network using a pair of intensity images and distance images to generate a Pix2Pix network optimization model, comprising the following steps: collecting vehicle data of a Gm-APD laser radar to provide the pair of intensity images and distance images; inputting the vehicle data into the Pix2Pix network to obtain the Pix2Pix network optimization model; an enhanced distance image acquisition unit, configured to input the enhanced intensity image into the Pix2Pix network optimization model to obtain an enhanced distance image corresponding to the enhanced intensity image. The Pix2Pix network structure mainly comprises two parts, a generator and a discriminator.

5. The deep learning-based Gm-APD lidar data augmentation system of claim 4, wherein, The generator G uses a Unet structure, and the input profile x is encoded and then decoded into a real picture. The discriminator D uses a conditional discriminator PatchGAN. The reinforced generative adversarial network acquisition unit comprises the following steps: ， wherein, denotes the mathematical expectation from real samples, denotes noise z the mathematical expectation from simulated samples generated by the generator, D ( x ) denotes x the probability that a sample is classified as real if it comes from real training data x , D ( x ) is considered true, outputting a score close to 1, otherwise it is considered x false, outputting a score close to 0, 1- D [ G ( z ) is the probability that a simulated sample G ( z ) is judged false.

6. The deep learning-based Gm-APD lidar data augmentation system of claim 4, wherein, training the generative adversarial network after adding the attention module, wherein the generative adversarial network after adding the attention module comprises a generative model G and a discriminative model D; training the generative model G and the discriminative model D, comprising the following steps: The enhanced distance image acquisition unit specifically comprises the following steps: ， wherein, is a loss function for the generator, is a real image, is a generated image, is a difference between the real image and the generated image.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium is used to store a computer program, and the computer program executes the deep learning-based Gm-APD lidar data enhancement method in any one of claims 1-3.

8. A computer device, comprising: The computer readable storage medium is used to store a computer program, and the computer program executes the deep learning-based Gm-APD lidar data enhancement method in any one of claims 1-3.

Images