Knowledge distillation-based vector map construction model training method and device

Abstract

The embodiment of the invention provides a vector map construction model training method and device based on knowledge distillation, and relates to an automatic driving perception technology method.The method comprises the steps that a teacher model is constructed by constructing a first vector map, and a student model is constructed by constructing a second vector map; determining Bayesian loss of the first vector map construction teacher model for training, determining intermediate layer feature loss and optimal weight loss of the trained first vector map construction teacher model and the second vector map construction student model, and determining first loss of each intermediate layer feature of the second vector map construction student model, and adjusting parameters of the second vector map construction student model based on the intermediate layer feature loss, the optimal weight loss and the first loss so as to train the second vector map construction student model, and constructing a map based on the second vector map construction student model. The cost of map construction can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving perception, and more specifically, to a method and device for training a vector map construction model based on knowledge distillation. Background Art

[0002] In modern map construction, especially in fields such as autonomous driving and robot navigation, complex deep learning models are usually used for high-precision map construction. These models can include convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc. These models can process large amounts of data, extract rich features, and generate highly accurate map representations. However, these heavyweight models also have their inherent disadvantages: 1. High computational resource consumption: Heavyweight models require powerful computational resources (such as high-performance GPUs, large amounts of memory, etc.) to run, which increases the hardware cost and energy consumption.

[0003] 2. High training difficulty: Due to the complexity of the model, a large amount of labeled data and a long training period are required during the training process, and it is easy to fall into local optimal solutions.

[0004] 3. Inference latency: In real-time application scenarios, the model needs to complete inference quickly, while the inference speed of heavyweight models is slow, which may lead to system response latency.

[0005] 4. Difficult to deploy: Heavyweight models are difficult to directly deploy on edge devices because edge devices usually have limited computing power and small memory.

[0006] Although heavyweight models have significant advantages in map construction, they also have obvious limitations. Therefore, researchers are constantly exploring new methods to balance the performance and efficiency of the models. There is an urgent need for a vector map training method that makes map construction not only more accurate but also easier to deploy and use. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and device for training a vector map construction model based on knowledge distillation, which can reduce the deployment difficulty and the cost of vector map construction.

[0008] To achieve the above purpose, the technical solutions adopted in the embodiments of the present application are as follows: In a first aspect, an embodiment of the present application provides a method for training a vector map construction model based on knowledge distillation, the method including: Constructing a first vector map construction teacher model; Construct a second vector map to build a student model, where the first vector map to build a teacher model is a heavyweight model, and the second vector map to build a student model is a lightweight map model; Determine the Bayesian loss of the first vector map to build a teacher model; Train the first vector model based on the Bayesian loss; Determine the intermediate layer feature loss and the optimal weight loss between the trained first vector map to build a teacher model and the second vector map to build a student model; Determine the first loss of each intermediate layer feature of the second vector map to build a student model; Based on the intermediate layer feature loss, the optimal weight loss, and the first loss, determine to adjust the parameters of the second vector map to build a student model to train the second vector map to build a student model; Build a map based on the trained second vector map to build a student model.

[0009] In an alternative embodiment, the step of determining the intermediate layer feature loss between the trained first vector map to build a teacher model and the second vector map to build a student model includes: Determine the first intermediate layer features of each intermediate layer of the trained first vector map to build a teacher model and the first final output feature output by the last layer of the trained first vector map to build a teacher model; Based on the first final output feature and each of the first intermediate layer features, obtain each enhanced feature; Determine the second intermediate layer features of each intermediate layer of the second vector map to build a student model; Align each of the second intermediate layer features with the corresponding enhanced feature to obtain the aligned second intermediate layer features; Calculate the standard cross-entropy loss of the second vector map to build a student model; Based on each of the enhanced features, each of the aligned second intermediate layer features, and the standard cross-entropy loss, obtain the intermediate layer feature loss between the trained first vector map to build a teacher model and the second vector map to build a student model.

[0010] In an alternative embodiment, the step of obtaining each enhanced feature based on the first final output feature and each of the first intermediate layer features includes: For each first intermediate layer feature, calculate the first product of the first intermediate layer feature and the first final output feature; Calculate the first sum of the first product and the first final output feature as the enhanced feature.

[0011] In an alternative embodiment, the intermediate layer feature loss satisfies the following formula: ; wherein, is the intermediate layer feature loss, is the standard cross-entropy loss, is the aligned second intermediate layer feature, is the enhanced feature.

[0012] In an alternative embodiment, the steps of determining the optimal weight loss of the trained first vector map construction teacher model and the second vector map construction student model include: Determine the first predicted pixel label of the vector map image output by the trained first vector map construction teacher model; Determine the second predicted pixel label of the vector map image output by the second vector map construction student model; Based on the first predicted pixel label and the second predicted pixel label, obtain the optimal weight loss of the trained first vector map construction teacher model and the second vector map construction student model.

[0013] In an alternative embodiment, the optimal weight loss satisfies the following formula: ; wherein, is the optimal weight loss, X is the first predicted pixel label, and Y is the second predicted pixel label.

[0014] In an alternative embodiment, the steps of determining the first loss of each intermediate layer feature of the second vector map construction student model include: Based on each of the enhanced features and each of the aligned second intermediate layer features, calculate the first loss of each intermediate layer feature of the second vector map construction student model; wherein, the first loss satisfies the following formula: ; wherein, is the first loss, is the aligned second intermediate layer feature, is the enhanced feature.

[0015] In an alternative embodiment, the steps of constructing the first vector map construction teacher model include: Construct the first vector map construction teacher model based on a frame-level map network, a temporal fusion module of Mamba, an inter-segment decoder, an intra-segment decoder, Mamba, and a DETR encoder; The frame-level map network is used to process the acquired multi-view image fragment set to obtain the BEV features and map queries of the multi-view image fragment set, where the multi-view image fragment set includes image fragments corresponding to different views; The temporal fusion module of the Mamba is used to fuse each of the BEV features with the map queries to obtain a plurality of fused BEV features; The inter-segment decoder is used to determine the segment markers of each image fragment based on each of the BEV features and the map queries; for each fused BEV feature, update the fused BEV feature based on the segment marker of the fused BEV feature and the segment markers of the image fragments other than the segment marker of the fused BEV feature to obtain an updated fused BEV feature; obtain map markers based on each of the map queries and each updated fused BEV feature; The inter-segment decoder is used to write the map markers and the segment markers of the fused BEV features into the memory; The Mamba and the DETR encoder generate a segment-level vector map to construct the first vector map construction teacher model.

[0016] In an alternative embodiment, the temporal fusion module of the Mamba is specifically used for: For each of the BEV features, obtain the historical BEV feature of the BEV feature; Align the BEV feature and the historical BEV feature corresponding to the BEV feature; Concatenate the aligned BEV feature and the historical BEV feature into a first BEV feature; Perform a first rule process and a second rule process on the first BEV feature respectively to obtain a first result and a second result, where the first rule process includes processing the first BEV feature sequentially through a 3×3 convolutional layer, a batch normalization layer, and a ReLU layer, and the second rule process includes processing the first BEV feature sequentially through a 1×1 convolutional layer and a batch normalization layer; Add the first result and the second result to obtain a third result; Compress the third result sequentially through a linear layer and a layer normalization layer to obtain a third compressed result; After rearranging the third compressed result, obtain a rearranged third compressed result; Process the rearranged third compressed result through the Mamba2 model to obtain a fourth result; Query and merge the fourth result in a preset manner to obtain a fifth result; Perform average pooling on the fifth result to obtain the fused feature of the BEV feature.

[0017] In a second aspect, an embodiment of the present application provides an apparatus for training a vector map construction model based on knowledge distillation. The apparatus includes: A construction module, configured to construct a first vector map construction teacher model; and construct a second vector map construction student model, where the first vector map construction teacher model is a heavyweight model, and the second vector map construction student model is a lightweight map model; A determination module, configured to determine the Bayesian loss of the first vector map construction teacher model; A first training module, configured to train the first vector model based on the Bayesian loss; The determination module is further configured to determine the intermediate layer feature loss and the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model; and determine the first loss of each intermediate layer feature of the second vector map construction student model; A parameter adjustment module, configured to determine, based on the intermediate layer feature loss, the optimal weight loss, and the first loss, to adjust the parameters of the second vector map construction student model, so as to train the second vector map construction student model; A map construction module, configured to construct a map based on the second vector map construction student model. In a third aspect, an embodiment of the present application provides an electronic device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps of the method for training a vector map construction model based on knowledge distillation are implemented.

[0018] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method for training a vector map construction model based on knowledge distillation are implemented.

[0019] The present application has the following beneficial effects: In this application, a teacher model is constructed by building a first vector map, and a student model is constructed by building a second vector map. Among them, the first vector map for constructing the teacher model is a heavyweight model, and the second vector map for constructing the student model is a lightweight map model. The Bayesian loss of the first vector map for constructing the teacher model is determined, and the first vector map for constructing the teacher model is trained based on the Bayesian loss. The intermediate layer feature loss and the optimal weight loss between the trained first vector map for constructing the teacher model and the second vector map for constructing the student model are determined. The first loss of each intermediate layer feature of the second vector map for constructing the student model is determined. The parameters of the second vector map for constructing the student model are adjusted based on the intermediate layer feature loss, the optimal weight loss, and the first loss to train the second vector map for constructing the student model. The map is constructed based on the trained second vector map for constructing the student model. The pre-trained first vector map for constructing the teacher model is used as the teacher model to guide the lightweight second vector map for constructing the student model to learn. The compression and acceleration of the model are achieved, making it more suitable for deployment and application in resource-constrained environments. At the same time, when constructing the map based on the trained second vector map for constructing the student model, the cost of map construction can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0021] Figure 1 It is a block diagram of an electronic device provided by an embodiment of the present invention; Figure 2 It is one of the flowcharts of a method for training a vector map construction model based on knowledge distillation provided by an embodiment of the present invention; Figure 3 It is another flowchart of a method for training a vector map construction model based on knowledge distillation provided by an embodiment of the present invention; Figure 4 It is yet another flowchart of a method for training a vector map construction model based on knowledge distillation provided by an embodiment of the present invention; Figure 5 It is a structural block diagram of a device for training a vector map construction model based on knowledge distillation provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations.

[0023] Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0024] It should be noted that like reference numerals and letters denote like items in the following figures. Therefore, once an item is defined in one figure, it does not require further definition and explanation in subsequent figures.

[0025] In the description of the present invention, it should be noted that if terms such as "upper", "lower", "inner", "outer", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, or the orientations or positional relationships in which the inventive product is customarily placed during use, it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention.

[0026] In addition, terms such as "first", "second", etc. are only used for descriptive distinction and should not be construed as indicating or implying relative importance.

[0027] In the description of the present application, it should also be noted that unless otherwise clearly defined and limited, the terms "set", "installed", "connected", "coupled" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the internal communication of two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to specific circumstances.

[0028] Through extensive research by the inventors, it has been found that in modern map construction, especially in fields such as autonomous driving and robot navigation, complex deep learning models are usually used for high-precision map construction. These models can include convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc. These models can process large amounts of data, extract rich features, and generate highly accurate map representations. However, these heavyweight models also have their inherent disadvantages: 1. High computational resource consumption: Heavyweight models require powerful computational resources (such as high-performance GPUs, large amounts of memory, etc.) to run, which increases the hardware cost and energy consumption.

[0029] 2. High training difficulty: Due to the complexity of the model, a large amount of labeled data and a long training cycle are required during the training process, and it is easy to fall into local optimal solutions.

[0030] 3. Inference latency: In real-time application scenarios, the model needs to complete inference quickly, while the inference speed of heavyweight models is slow, which may lead to system response latency.

[0031] 4. Difficult to deploy: Heavyweight models are difficult to directly deploy on edge devices because edge devices usually have limited computing power and small memory.

[0032] Although heavyweight models have significant advantages in map construction, they also have obvious limitations. Therefore, researchers are constantly exploring new methods to balance the performance and efficiency of the models. There is an urgent need for a vector map training method that makes map construction not only more accurate but also easier to deploy and use.

[0033] In addition, in the existing technology, SLAM (Simultaneous Localization and Mapping) technology is relied on in multi-view high-precision map construction. SLAM technology is the main means of constructing static maps. Through steps such as feature extraction, matching, and mapping, SLAM technology can effectively construct a map of the environment and determine its own position. However, these methods are usually computationally complex, have a large amount of data processing, and are mainly applicable to map construction work in an offline environment.

[0034] In recent years, with the progress of camera technology, camera-based multi-view methods have gradually become a hot topic in constructing high-precision maps. These methods capture multi-perspective images, convert them into BEV (Bird's Eye View) features, and then use a specific decoder to predict map elements. This method has received extensive attention in practical applications due to its high cost-effectiveness and good real-time performance. However, these methods perform poorly in dealing with the problem of dynamic object occlusion, resulting in the accuracy of the prediction results being affected.

[0035] To improve prediction accuracy, researchers have begun to explore the use of temporal information to enhance the quality of map construction by analyzing information in consecutive frames. Although this direction shows good potential, existing methods still have deficiencies in processing inter-frame information over long time spans, especially in long-range perception and heavy occlusion scenarios, where performance is limited.

[0036] In view of the discovery of the above problems, this embodiment provides a method and apparatus for training a vector map construction model based on knowledge distillation. The method can construct a first vector map construction teacher model and a second vector map construction student model. Among them, the first vector map construction teacher model is a heavyweight model, and the second vector map construction student model is a lightweight map model. Determine the Bayesian loss of the first vector map construction teacher model, train the first vector map construction teacher model based on the Bayesian loss, determine the intermediate layer feature loss and optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model, determine the first loss of each intermediate layer feature of the second vector map construction student model, and determine the adjustment of the parameters of the second vector map construction student model based on the intermediate layer feature loss, optimal weight loss, and the first loss to train the second vector map construction student model, and construct a map based on the trained second vector map construction student model. Use the pre-trained first vector map construction teacher model as a teacher model to guide the lightweight second vector map construction student model to learn. It realizes the compression and acceleration of the model, making it more suitable for deployment and application in resource-constrained environments. At the same time, when constructing a map based on the trained second vector map construction student model, the cost of map construction can be reduced. The solution provided in this embodiment will be elaborated in detail below.

[0037] This embodiment provides an electronic device that can train a vector map construction model. In a possible implementation, the electronic device can be a user terminal. For example, the electronic device can be, but is not limited to, a server, a smart phone, a personal computer (PC), a tablet computer, a personal digital assistant (PDA), a mobile internet device (MID), etc.

[0038] Please refer to Figure 1 , Figure 1 which is a schematic structural diagram of the electronic device 100 provided in an embodiment of the present application. The electronic device 100 may further include more or fewer components than those shown in Figure 1 or have a different configuration from that shown in Figure 1 . Figure 1Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0039] The electronic device 100 includes a vector map construction model training device 110 based on knowledge distillation, a memory 120 and a processor 130.

[0040] The components of the memory 120 and the processor 130 are electrically connected to each other directly or indirectly to realize data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The vector map construction model training device 110 based on knowledge distillation includes at least one software function module that can be stored in the memory 120 in the form of software or firmware or solidified in the operating system (OS) of the electronic device 100. The processor 130 is used to execute the executable modules stored in the memory 120, such as the software function modules and computer programs included in the vector map construction model training device 110 based on knowledge distillation.

[0041] The memory 120 may be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc. The memory 120 is used to store a program, and the processor 130 executes the program after receiving an execution instruction.

[0042] Please refer to Figure 2 , Figure 2 For application Figure 1 A flowchart of a vector map construction model training method based on knowledge distillation for an electronic device 100 is provided, and the method including each step is described in detail below.

[0043] S201: Construct a first vector map to construct a teacher model.

[0044] S202: Construct a second vector map to construct a student model.

[0045] Among them, the first vector map constructs the teacher model as a heavyweight model, and the second vector map constructs the student model as a lightweight map model.

[0046] S203: Determine the Bayesian loss of the first vector map construction teacher model.

[0047] S204: Train the first vector map construction teacher model based on the Bayesian loss.

[0048] S205: Determine the intermediate layer feature loss and optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model.

[0049] S206: Determine the first loss of each intermediate layer feature of the second vector map construction student model.

[0050] S207: Adjust the parameters of the second vector map construction student model based on the intermediate layer feature loss, optimal weight loss, and the first loss to train the second vector map construction student model.

[0051] S208: Construct a map based on the trained second vector map construction student model.

[0052] It should be noted that the first vector map construction teacher model and the second vector map construction student model can be segment-level vector map construction models.

[0053] The segment-level vector map construction model is an algorithm architecture used to generate accurate and structured vector maps from local or segment data.

[0054] The lightweight model aims to reduce the number of parameters and computational complexity while maintaining relatively high performance as much as possible, and is suitable for resource-constrained devices (such as mobile devices, embedded systems). The characteristics of the lightweight model are: fewer parameters, usually only a few million to tens of millions of parameters, low computational volume, small memory footprint, and fast inference speed.

[0055] The heavyweight model usually has a large number of parameters and high computational complexity, can capture more complex features and patterns, and is suitable for high-performance computing devices (such as GPU / TPU clusters). The characteristics of the heavyweight model are: large number of parameters: usually hundreds of millions to tens of billions of parameters, high computational volume, large memory footprint, and excellent performance.

[0056] Vector maps use geometric objects (such as points, lines, polygons) to represent geospatial information.

[0057] In vector map construction, the Bayesian loss is a loss function designed through the Bayesian inference framework, aiming to combine the uncertainty of prior knowledge (such as geometric and topological constraints of the map) and observed data (such as sensor inputs) to optimize the parameters of the map model. Its core idea is to minimize the negative log-likelihood of the posterior distribution, thus transforming the map construction problem into an optimization problem in the probability framework.

[0058] Exemplarily, the Bayesian loss of the first vector map construction teacher model can obtain the image to be detected, determine the predicted value and the true value of each pixel of the image to be detected, and calculate the difference between the predicted value and the true value of each pixel point as the Bayesian loss of the first vector map construction teacher model.

[0059] Specifically, the Bayesian loss of the first vector map construction teacher model is calculated by the following formula: ; where represents the distance function, is the Bayesian loss using the posterior label probability and the estimated density map.

[0060] Based on the Bayesian loss, the first vector map construction teacher model is trained. When the Bayesian loss is greater than the preset value, the parameters of the first vector map construction teacher model are adjusted by the Bayesian loss to obtain the adjusted first vector map construction teacher model. Then, the Bayesian loss of the adjusted first vector map construction teacher model is determined again. When the Bayesian loss function is greater than the preset value, the parameters of the adjusted first vector map construction teacher model are adjusted based on the Bayesian loss to obtain a new adjusted first vector map construction teacher model. The new Bayesian loss is determined again until the new Bayesian loss converges, and then the training of the first vector map construction teacher model is completed.

[0061] Based on the trained first vector map construction teacher model, the training of the second vector map construction student model is guided, so that the second vector map construction student model can have the prediction effect of the trained first vector map construction teacher model.

[0062] Specifically, by determining the intermediate layer feature loss and the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model. In addition, the second vector map construction student model performs self-training based on the intermediate layer features, and determines the first loss of each intermediate layer feature of the second vector map construction student model. The second vector map construction student model is trained based on the above multiple losses, and the trained second vector map construction student model is used for map construction.

[0063] The intermediate layer features of the model refer to the feature vectors or tensors generated by the hidden layer inside the neural network or other machine learning models. These features are obtained by gradually applying a series of transformations (such as convolution, activation function, pooling, etc.) to the input data.

[0064] The purpose of determining the intermediate layer feature loss between the trained first vector map construction teacher model and the second vector map construction student model is as follows: to transfer the rich feature information of the intermediate layer of the trained first vector map construction teacher model to the second vector map construction student model, so as to enrich the feature representation of the second vector map construction student model and enable it to obtain performance similar to that of the trained first vector map construction teacher model while maintaining light weight.

[0065] The purpose of determining the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model is as follows: to ensure that the final output features of the second vector map construction student model are highly consistent with the features of the trained first vector map construction teacher model globally, which helps the second vector map construction student model to better capture the global feature information during the final prediction.

[0066] Determine the first loss of each intermediate layer feature in the second vector map construction student model, so that the second vector map construction student model not only obtains knowledge from the trained first vector map construction teacher model during the learning process, but also optimizes its own performance through its own intermediate layer features.

[0067] Finally, based on the intermediate layer feature loss, the optimal weight loss, and the first loss, determine the adjustment of the parameters of the first vector map construction teacher model and the second vector map construction student model until the sum of the final intermediate layer feature loss, the optimal weight loss, and the first loss converges, that is, the training of the second vector map construction student model is completed.

[0068] When adjusting the parameters of the second vector map construction student model based on the intermediate layer feature loss, the optimal weight loss, and the first loss, it is possible to determine the sum of the intermediate layer feature loss, the optimal weight loss, and the first loss, and adjust the parameters of the second vector map construction student model. When the sum of the intermediate layer feature loss, the optimal weight loss, and the first loss does not converge, then re-determine the sum of the intermediate layer feature loss, the optimal weight loss, and the first loss, adjust the parameters of the second vector map construction student model after adjusting the parameters, and determine the sum of the intermediate layer feature loss, the optimal weight loss, and the first loss between the trained first vector map construction teacher model and the second vector map model after adjusting the parameters again. Repeat the iteration when it does not converge until the sum of the intermediate layer feature loss, the optimal weight loss, and the first loss converges, that is, the training of the second vector map construction student model is completed.

[0069] Finally, perform map construction based on the trained second vector map construction student model, so that when constructing the model, using the trained vector map construction model reduces the training cost, and using the trained second vector map construction student model is more suitable for deployment and application in resource-constrained environments.

[0070] There are various ways to implement the intermediate layer feature loss between the trained first vector map construction teacher model and the second vector map construction student model. In one implementation, as Figure 3 shown, the following steps are included: S301: Determine the first intermediate layer features of each intermediate layer of the trained first vector map construction teacher model and the first final output feature output by the last layer of the trained first vector map construction teacher model.

[0071] It should be noted that the trained first vector map construction teacher model includes multiple layers. Exemplarily, when the trained first vector map construction teacher model includes the first layer, the second layer, the third layer, and the fourth layer, when the image to be detected is input into the trained first vector map construction teacher model, the outputs of the first layer, the second layer, and the third layer are extracted as intermediate layer features, and the output of the fourth layer is extracted as the first final output feature.

[0072] S302: Obtain each enhanced feature based on the first final output feature and each first intermediate layer feature.

[0073] There are various ways to obtain each enhanced feature based on the first final output feature and each first intermediate layer feature. In one implementation, for each first intermediate layer feature, calculate the first product of the first intermediate layer feature and the first final output feature, and calculate the first sum of the first product and the first final output feature as the enhanced feature.

[0074] Exemplarily, the intermediate layer features of the trained first vector map construction teacher model can be processed through a consistency enhancement module. The consistency enhancement module multiplies the first final output feature of the trained first vector map construction teacher model by each intermediate layer feature, and finally obtains each enhanced feature. Each enhanced feature highlights the common information more prominently, and the enhanced features retain more original excellent features through the way of residual connection.

[0075] S303: Determine the second intermediate layer features of each intermediate layer of the second vector map construction student model.

[0076] The method of determining the second intermediate features of each intermediate layer in the first vector map construction teacher model is similar to the method of determining the first intermediate layer features of each intermediate layer in the trained first vector map construction teacher model, and will not be elaborated here.

[0077] S304: Align each second intermediate layer feature with the corresponding enhanced feature to obtain the aligned second intermediate layer features.

[0078] There are multiple ways to align each second intermediate layer feature with the corresponding enhanced feature. In one implementation: The second intermediate layer features of the second vector map construction student model can be adjusted to the same resolution as the enhanced features of the trained first vector map construction teacher model through convolution, upsampling, and normalization layers, so that features at different levels can be better aligned and fused.

[0079] S305: Calculate the standard cross-entropy loss of the second vector map construction student model.

[0080] The standard cross-entropy loss is the difference between the probability distribution output by the second vector map construction student model and the true label distribution.

[0081] S306: Based on each enhanced feature, each aligned second intermediate layer feature, and the standard cross-entropy loss, obtain the intermediate layer feature loss between the trained first vector map construction teacher model and the second vector map construction student model.

[0082] It should be noted that the intermediate layer feature loss can use the AT loss with a relatively fast convergence rate as the distillation function.

[0083] Specifically, the intermediate layer feature loss satisfies the following formula: ; Where is the intermediate layer feature loss, is the standard cross-entropy loss, is the aligned second intermediate layer feature, is the enhanced feature.

[0084] There are multiple ways to determine the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model. In one implementation, as Figure 4 shown, it includes the following steps: S401: Determine the first predicted pixel label of the vector map image output by the trained first vector map construction teacher model.

[0085] S402: Determine the second predicted pixel label of the vector map image output by the second vector map construction student model.

[0086] S403: Based on the first predicted pixel label and the second predicted pixel label, obtain the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model.

[0087] Exemplarily, the image to be detected is input into the trained first vector map construction teacher model. The trained first vector map construction teacher model outputs a segment-level vector map. Each frame of the image in the segment-level vector map is determined. The pixels in each frame of the image and the corresponding first predicted pixel labels are determined. The same image to be detected is input into the second vector map construction student model. The second vector map construction student model outputs a segment-level vector map. Each frame of the image in the segment-level vector map is determined. The pixels in each frame of the image and the corresponding second predicted pixel labels are determined. The difference between the first predicted pixel label and the second predicted pixel label is calculated as the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model.

[0088] The purpose of calculating the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model is to transfer the optimal weights and global feature representations of the trained first vector map construction teacher model to the second vector map construction student model, thereby improving the performance of the second vector map construction student model. This training method not only enriches the global feature representations of the second vector map construction student model but also optimizes the final output, improving the prediction accuracy and generalization ability.

[0089] Before calculating the optimal weight loss, the vector map image output by the second vector map construction student model can be passed through convolutional, upsampling, and normalization layers to adjust the features of the second vector map construction student model to the same resolution as the first vector map features, ensuring alignment in both spatial and channel dimensions.

[0090] Specifically, the optimal weight loss satisfies the following formula: ; where is the optimal weight loss, X is the first predicted pixel label, and Y is the second predicted pixel label.

[0091] There are multiple ways to determine the first loss of the features of each intermediate layer of the second vector map construction student model. In one implementation: Based on each enhanced feature and each of the aligned second intermediate layer features, the first loss of the features of each intermediate layer of the second vector map construction student model is calculated.

[0092] Specifically, the first loss satisfies the following formula: ; where is the first loss, is the aligned second intermediate layer feature, is the enhanced feature.

[0093] The student model for constructing the second vector map is tuned and trained based on the first loss, so that the student model for constructing the second vector map not only obtains knowledge from the trained teacher model for constructing the first vector map during the learning process, but also optimizes its own performance through its aligned second intermediate layer features. This not only improves the quality of the feature representation of the student model for constructing the second vector map, but also reduces redundant information, improving performance and generalization ability.

[0094] Determining the first loss of the intermediate layer features of the student model for constructing the second vector map is implemented in multiple stages. In each stage, the student model for constructing the second vector map optimizes according to its own intermediate layer features, gradually improving its own performance. This multi-stage implementation method enables the student model for constructing the second vector map to gradually learn more complex feature representations.

[0095] Exemplarily, the KLD loss function is used to quantify the differences between the intermediate layer features of the student model for constructing the second vector map. The KLD loss function can measure the similarity between two probability distributions and is used to evaluate the differences in the feature representations of the student model for constructing the second vector map at different stages.

[0096] Through the KLD loss function, the student model for constructing the second vector map will perform self-sublimation in each stage, that is, optimize its own feature representation. This enables the student model for constructing the second vector map to continuously improve its own performance during the learning process and finally reach a better state.

[0097] There are multiple implementation methods for constructing the teacher model for constructing the first vector map. In one implementation method: The teacher model for constructing the first vector map is constructed based on a frame-level map network, the temporal fusion module of Mamba, an inter-segment decoder, an intra-segment decoder, Mamba, and a DETR encoder.

[0098] In the construction of multi-view high-precision maps, traditional methods rely on SLAM technology to construct static maps through feature extraction, matching, and mapping. However, these methods are computationally complex and are only applicable to offline applications. In recent years, camera-based multi-view methods have received attention due to their cost-effectiveness and real-time performance. These methods convert multi-perspective images into BEV features and then use specific decoders to predict map elements, but are often affected by dynamic object occlusion, which affects prediction accuracy.

[0099] Therefore, to solve the above problems, when constructing the teacher model for constructing the first vector map in this application, a frame-level map network, the temporal fusion module of Mamba, an inter-segment decoder, an intra-segment decoder, Mamba, and a DETR encoder are used for construction.

[0100] Therefore, by constructing the first vector map construction teacher model based on the frame-level map network, the temporal fusion module of Mamba, the inter-segment decoder, the intra-segment decoder, Mamba, and the DETR encoder, the accuracy of map construction can be improved, as well as the efficiency of map construction.

[0101] The frame-level map network is used to process the obtained multi-view image segment set to obtain the BEV features and map queries of the multi-view image segment set, where the multi-view image segment set includes image segments corresponding to different views. The temporal fusion module of Mamba is used to fuse each BEV feature with the map query to obtain multiple fused BEV features.

[0102] The temporal fusion module of Mamba is specifically used for: for each BEV feature, obtaining the historical BEV feature of the BEV feature; aligning the BEV feature with the corresponding historical BEV feature; splicing the aligned BEV feature and historical BEV feature into the first BEV feature; respectively performing the first rule processing and the second rule processing on the first BEV feature to obtain the first result and the second result, where the first rule processing includes processing the first BEV feature through a 3×3 convolutional layer, a batch normalization layer, and a ReLU layer in sequence, and the second rule processing includes processing the first BEV feature through a 1×1 convolutional layer and a batch normalization layer in sequence; adding the first result and the second result to obtain the third result; compressing the third result through a linear layer and a layer normalization layer in sequence to obtain the third compressed result; rearranging the third compressed result to obtain the rearranged third compressed result; processing the rearranged third compressed result through the Mamba2 model to obtain the fourth result; querying and merging the fourth result in a preset manner to obtain the fifth result; performing average pooling on the fifth result to obtain the fused feature of the BEV feature.

[0103] Specifically, based on the temporal fusion module of Mamba, for each BEV feature, the historical BEV feature of the BEV feature is first obtained. Exemplarily, in the case of including BEV feature 1, BEV feature 2, BEV feature 3, BEV feature 4, and BEV feature 5, when fusing BEV feature 5, BEV feature 1, BEV feature 2, BEV feature 3, and BEV feature 4 are the historical BEV features of BEV feature 5.

[0104] There are various ways to align the BEV feature with the corresponding historical BEV feature. The alignment method can be alignment based on the attention mechanism, alignment based on the loss function, alignment based on the adversarial network, alignment based on statistics, etc. The embodiments of the present application do not make specific limitations on this.

[0105] The aligned historical BEV feature is concatenated with the BEV feature. Exemplarily, for the fusion of BEV feature 2, BEV feature 1 is the historical BEV feature of BEV feature 2. After aligning BEV feature 1 and BEV feature 2, the dimension of the aligned BEV feature 1 is 256, and the dimension of the aligned BEV feature 2 is 256. The aligned BEV feature 1 and BEV feature 2 are concatenated to form a new BEV feature, which is the first BEV feature with a dimension of 512.

[0106] The first BEV feature is processed using different rule-based processing methods to obtain different processing results. Exemplarily, the first BEV feature is processed through a series of 3×3 convolutional layers, batch normalization layers, and ReLU activation functions to obtain the first result. The first BEV feature is processed through a 1×1 convolutional layer and a batch normalization layer to obtain the second result. Processing through 3×3 convolutional layers, batch normalization layers, and ReLU activation functions is used to process the local detail features in the image, and processing through a 1×1 convolutional layer and a batch normalization layer is used to process the global features in the image. The processing results of the two methods, namely the first result and the second result, are added to obtain the third result, and the added result is successively compressed through a linear layer and a layer normalization layer to obtain a feature map with a dimension of 256, which is the third compression result. The main role of the linear layer is to project the added result into a low-dimensional space to achieve feature compression. The layer normalization layer normalizes the compressed features to ensure the stability of the training process.

[0107] The third compression result is rearranged. The rearrangement method can include a multi-directional feature sequence scanning mechanism, such as rearrangement in four directions: forward left, forward up, backward left, and backward up. This multi-directional rearrangement method avoids the disadvantages of the serpentine spiral recombination method and prevents the loss of interaction between adjacent features.

[0108] The rearranged third compression result is processed through the Mamba2 model to obtain the fourth result. The feature sequence of the rearranged third compression result is input into the Mamba2 block for processing. The Mamba2 block utilizes its improved structural state space duality (SSD) theoretical framework and combines the self-attention mechanism to enhance the utilization of temporal information and obtains the fourth result with enhanced temporal information utilization. The output of the Mamba2 block is recombined into the original feature map order. To further enhance the effect of feature fusion, the output feature map is added to the current feature map through skip connections, and a dropout operation may be applied. The output of the Mamba2 block also undergoes a query merging operation to generate the final feature representation, which is the fifth result. The fifth result is a feature map with a dimension of 256. After average pooling processing of the fifth result, the fused feature of the BEV feature is obtained.

[0109] The temporal fusion module of Mamba fuses each BEV feature with the map query, and the implementation method of obtaining multiple fused BEV features satisfies the following formula:

[0110] Among them, represents the current BEV feature, represents the historical BEV feature, represents the first result, represents the second result, represents the third result after being compressed by a linear layer and a layer normalization layer, represents the processing of a convolutional layer, represents the processing of a batch normalization layer, represents the processing of a layer normalization layer, represents the processing of a linear layer.

[0111] Mamba is a new type of sequence processing model designed to address the computational complexity issues faced by traditional Transformers when dealing with long sequences. It is based on a structured state space model and improves the training speed through block-decomposed matrix multiplication and GPU memory hierarchy optimization, and shows performance comparable to Transformers in multiple downstream tasks. As an improved version of Mamba, Mamba2 further introduces architectural modifications such as parallel projection, additional normalization layers, and grouped head patterns, enhancing performance and scalability.

[0112] The inter-segment decoder is used to determine the segment tokens of each image segment based on each BEV feature and the map query; for each fused BEV feature, the fused BEV feature is updated based on the segment tokens of the fused BEV feature and the segment tokens of each image segment other than the segment tokens of the fused BEV feature to obtain the updated fused BEV feature.

[0113] The inter-segment decoder processes the information between different input image segments, including read operations and write operations.

[0114] The inter-segment decoder reads the segment tokens of the multi-view image segment set from the memory through read operations, and generates the segment tokens of the current segment in combination with the map query output by the frame-level map network.

[0115] Exemplarily, in the output of the Mamba2 block, features related to segment tokens are extracted through a Query mechanism. For example, in an object detection task, a class token can be used to represent the semantic information of the entire sequence. The output of the Mamba2 block is recombined into a feature map, and the final segment tokens are generated through a task head (such as a segmentation head or a multi-layer perceptron MLP). The output features are normalized and input into a prediction head to generate semantic labels or object detection results for each segment.

[0116] The BEV feature updater updates the fused BEV feature based on the fused BEV feature output by the Mamba-based temporal fusion module, the segment tokens of the current segment, and the segment tokens of other segments.

[0117] BEV features may not directly reveal all hidden map element information, such as occluded objects or dynamically changing environments. By combining temporal information and the token information of other segments, the BEV feature updater can enhance the perception ability of these hidden information. Segment tokens provide semantic information to help the model understand the key regions in the BEV features. Embedding the token information of the current segment into the BEV feature enhances the semantic understanding of the current segment, and using the token information of other segments provides context support to help the model understand the environment of the current segment. For example, through an attention mechanism or feature fusion, the token information of other segments is interacted with the BEV feature of the current segment. The fused BEV feature, the segment tokens of the current segment, and the segment tokens of other segments are fused through a convolutional layer or MLP, and an attention module is used to highlight important feature regions and suppress irrelevant noise. The updated features are combined with the original BEV features through skip connections or residual connections to enhance the representation ability of the features. The updated BEV feature can better reflect the dynamic and static information in the scene, thereby revealing hidden map elements.

[0118] The map token generator of the intra-segment decoder generates map tokens based on the map queries generated by the frame-level map network and the updated BEV features. The inter-segment decoder writes the map tokens and the segment tokens of the current segment into the memory for system use through a write operation.

[0119] Map tokens are obtained based on each map query and each updated fused BEV feature; the inter-segment decoder is used to write the map tokens and the segment tokens of the fused BEV feature into the memory; Mamba and the DETR encoder generate a segment-level vector map to construct a first vector map construction teacher model.

[0120] Please refer to Figure 5 , this embodiment of the present application also provides an application for Figure 1The training device 110 for the vector map construction model based on knowledge distillation of the electronic device 100, the training device 110 for the vector map construction model based on knowledge distillation includes: A construction module 111, configured to construct a first vector map construction teacher model; construct a second vector map construction student model, wherein, the first vector map construction teacher model is a heavyweight model, and the second vector map construction student model is a lightweight map model; A determination module 112, configured to determine the Bayesian loss of the first vector map construction teacher model; A first training module 113, configured to train the first vector map construction teacher model based on the Bayesian loss; The determination module 112 is further configured to determine the intermediate layer feature loss and the optimal weight loss between the trained first vector map construction teacher model and the second vector map construction student model; determine the first loss of each intermediate layer feature of the second vector map construction student model; A parameter tuning module 114, configured to adjust the parameters of the second vector map construction student model based on the intermediate layer feature loss, the optimal weight loss, and the first loss, so as to train the second vector map construction student model; A map construction module 115, configured to construct a map based on the second vector map construction student model.

[0121] This application further provides an electronic device 100, and the electronic device 100 includes a processor 130 and a memory 120. The memory 120 stores computer-executable instructions, and when the computer-executable instructions are executed by the processor 130, the method for training the vector map construction model based on knowledge distillation is implemented.

[0122] This application embodiment further provides a computer-readable storage medium, and the storage medium stores a computer program, and when the computer program is executed by the processor 130, the method for training the vector map construction model based on knowledge distillation is implemented.

[0123] In the embodiments provided in this application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of devices, methods, and computer program products according to multiple embodiments of this application. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0124] In addition, each functional module in various embodiments of this application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part. If the function is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0125] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the said element.

[0126] As described above, the above are only various embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A vector map construction model training method based on knowledge distillation, characterized in that: The method comprises: Construct the first vector map to construct the teacher model; Constructing a second vector map construction student model, wherein the first vector map construction teacher model is a heavyweight model and the second vector map construction student model is a lightweight map model; determining a Bayesian loss for constructing a teacher model for the first vector map; Training the first vector map construction teacher model based on the Bayesian loss; Determine the intermediate layer feature loss and optimal weight loss of the trained first vector map construction teacher model and the second vector map construction student model; Determining a first loss for each intermediate layer feature of the student model constructed by the second vector map; Adjusting parameters of the second vector map construction student model based on the intermediate layer feature loss, the optimal weight loss, and the first loss to train the second vector map construction student model; A student model is constructed based on the second vector map to construct a map.

2. The method according to claim 1, characterized in that The step of determining the intermediate layer feature loss of the trained first vector map to construct the teacher model and the second vector map to construct the student model comprises: Determine first intermediate layer features of each intermediate layer of the trained first vector map construction teacher model and first final output features of the last layer output of the trained first vector map construction teacher model; Obtaining enhanced features based on the first final output features and the first intermediate layer features; determining second intermediate layer features of each intermediate layer of the student model constructed by the second vector map; Aligning each of the second intermediate layer features with the corresponding enhanced features to obtain aligned second intermediate layer features; Calculate the standard cross entropy loss of the student model constructed by the second vector map; Based on each of the enhanced features, each of the aligned second intermediate layer features and the standard cross entropy loss, the intermediate layer feature loss of the trained first vector map construction teacher model and the second vector map construction student model is obtained.

3. The method according to claim 2, characterized in that The step of obtaining each enhanced feature based on the first final output feature and each first intermediate layer feature comprises: For each first intermediate layer feature, calculating a first product of the first intermediate layer feature and the first final output feature; A first sum of the first product and the first final output feature is calculated as an enhanced feature.

4. The method according to claim 2, characterized in that: The intermediate layer feature loss satisfies the following formula: ; in, is the feature loss of the middle layer, is the standard cross entropy loss, is the second middle layer feature after alignment, To enhance the features.

5. The method according to claim 3, characterized in that: The step of determining the optimal weight loss of the trained first vector map construction teacher model and the second vector map construction student model comprises: Determining a first predicted pixel label of a vector map image output by the first trained vector map construction teacher model; determining a second predicted pixel label of the vector map image output by the second vector map construction student model; Based on the first predicted pixel label and the second predicted pixel label, an optimal weight loss of the trained first vector map construction teacher model and the second vector map construction student model is obtained.

6. The method according to claim 5, characterized in that The optimal weight loss satisfies the following formula: ； in, is the optimal weight loss, X is the first predicted pixel label, and Y is the second predicted pixel label.

7. The method according to claim 2, characterized in that The step of determining the first loss of each intermediate layer feature of the student model constructed by the second vector map comprises: Based on each of the enhanced features and each of the aligned second intermediate layer features, calculating the first loss of each intermediate layer feature of the student model constructed by the second vector map; The first loss satisfies the following formula: ； in, For the first loss, is the second middle layer feature after alignment, To enhance the features.

8. The method according to claim 1, characterized in that The step of constructing the first vector map to construct the teacher model includes: The first vector map construction teacher model is constructed based on the frame-level map network, Mamba's temporal fusion module, inter-segment decoder, intra-segment decoder, Mamba and DETR encoder; The frame-level map network is used to process the acquired multi-view image fragment set to obtain BEV features and map queries of the multi-view image fragment set, wherein the multi-view image fragment set includes image fragments corresponding to different viewpoints; The temporal fusion module of Mamba is used to fuse each of the BEV features with the map query to obtain a plurality of fused BEV features; The inter-segment decoder is used to determine a segment tag of each image segment based on each of the BEV features and the map query; for each fused BEV feature, based on the segment tag of the fused BEV feature and the segment tags of each image segment other than the segment tag of the fused BEV feature, update the fused BEV feature to obtain an updated fused BEV feature; and obtain a map tag based on each of the map queries and each updated fused BEV feature; The inter-segment decoder is used to write the map mark and the segment mark of the fused BEV feature into a memory; The Mamba and the DETR encoder generate segment-level vector maps to construct the first vector map to construct the teacher model.

9. The method according to claim 8, characterized in that The time fusion module of Mamba is specifically used for: For each of the BEV features, obtaining a historical BEV feature of the BEV feature; Aligning the BEV feature with a historical BEV feature corresponding to the BEV feature; splicing the aligned BEV feature and the historical BEV feature into a first BEV feature; Performing a first rule processing and a second rule processing on the first BEV feature respectively to obtain a first result and a second result, wherein the first rule processing includes sequentially passing the first BEV feature through a 3×3 convolution layer, a batch normalization layer, and a ReLU layer, and the second rule processing includes sequentially passing the first BEV feature through a 1×1 convolution layer and a batch normalization layer; Add the first result and the second result to obtain a third result; Compressing the third result through a linear layer and a layer normalization layer in sequence to obtain a third compressed result; After rearranging the third compression result, a rearranged third compression result is obtained; Processing the rearranged third compression result through the Mamba2 model to obtain a fourth result; The fourth result is searched in a preset manner and combined to obtain a fifth result; After average pooling is performed on the fifth result, a fusion feature of the BEV feature is obtained.

10. A vector map construction model training device based on knowledge distillation, characterized in that: The device comprises: A construction module, used to construct a first vector map construction teacher model; construct a second vector map construction student model, wherein the first vector map construction teacher model is a heavyweight model, and the second vector map construction student model is a lightweight map model; A determination module, configured to determine the Bayesian loss of the teacher model constructed by the first vector map; A first training module, configured to train the first vector map construction teacher model based on the Bayesian loss; The determination module is also used to determine the intermediate layer feature loss and optimal weight loss of the first vector map constructed teacher model and the second vector map constructed student model that have been trained; determine the first loss of each intermediate layer feature of the second vector map constructed student model; A parameter adjustment module, used for adjusting the parameters of the second vector map construction student model based on the intermediate layer feature loss, the optimal weight loss and the first loss, so as to train the second vector map construction student model; A map construction module is used to construct a student model based on the second vector map to construct a map.

Images