Vehicle-mounted laser point cloud target semantic domain adaptive semantic segmentation method
By decoupling the point cloud semantic segmentation model and using pseudo-labels and cross-coupled loss functions, the catastrophic forgetting problem in incremental learning is solved, and the model's recognition ability of old classes and new categories is achieved.
Patent Information
- Application Number
- CN202411453640.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-17
- Publication Date
- 2025-02-07
AI Technical Summary
The existing technology has catastrophic forgetting problems in point cloud semantic segmentation incremental learning, resulting in a significant decline in the recognition ability of the pre-trained model for old classes after class increment.
By decoupling the point cloud semantic segmentation model into feature encoder and classifier, the pre-trained model generates pseudo-labels for the target semantic space point cloud, and a cross-coupled loss function is constructed to realize that the model does not forget the old class knowledge while performing it while doing it. New category learning.
It effectively enhances the universality of the point cloud semantic segmentation increment method, ensuring that the model can still maintain the ability to identify old classes after class increment, and at the same time learn new categories.
Smart Images
Figure CN119399463A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of point cloud deep learning and incremental learning, and in particular to a vehicle-mounted laser point cloud target semantic domain adaptive semantic segmentation method. Background Art
[0002] Semantic Segmentation of 3D Point Cloud plays an increasingly important role with the rise of high-tech technologies such as vehicle-mounted laser scanning systems. The training of traditional semantic segmentation models requires the collection of required point cloud data in advance and labeling them. As a result, there are still many challenges in updating the model efficiently and intelligently for the recognition and classification of new targets.
[0003] In recent years, with the development of deep learning of point clouds, semantic segmentation of point clouds based on deep learning has also made great breakthroughs. Semantic segmentation of point clouds requires labeling a semantic category for each point in the point cloud data. However, since point cloud data often occupies a lot of memory resources, or due to copyright issues, the original point cloud data cannot be obtained during the incremental learning process. If the labels of the old categories in the newly added point cloud data are manually labeled again, it often wastes a lot of resources and time.
[0004] Compared with traditional point cloud semantic segmentation training methods, adding category recognition capabilities to the original semantic segmentation model can reduce the cost of training new models and have better flexibility. As a data-driven method, deep learning itself not only relies on a large amount of training data, but is also very sensitive to the parameters of the trained model. During the incremental learning process, the perturbation of the pre-trained semantic segmentation model may have a very serious impact, that is, the original model's recognition ability for old categories is greatly reduced, which is called catastrophic forgetting. As a result, the target semantic segmentation model obtained after the category increment of the pre-trained semantic segmentation model cannot meet the requirements.
[0005] However, there is currently no mature algorithm at home and abroad that can solve the problem of catastrophic forgetting in incremental learning of point cloud semantic segmentation. Summary of the invention
[0006] The purpose of the present invention is to propose a semantic domain adaptive semantic segmentation method for vehicle-mounted laser point cloud targets, so as to solve the application generalization problem of cross-lidar configurations of the deep learning-based vehicle-mounted laser point cloud semantic segmentation method for practical applications.
[0007] To achieve the above objectives, the present invention adopts the following technical solutions.
[0008] A semantic domain adaptive semantic segmentation method for vehicle-mounted laser point cloud targets can be based on a pre-trained point cloud semantic segmentation model, and in the case of not having original data but only having data of newly added classes, the category cognition of the pre-trained point cloud semantic segmentation model can be expanded, including the following steps;
[0009] Step S1. This project first divides the original open source dataset into two parts by combining the existing vehicle-mounted laser point cloud semantic segmentation open source dataset. One part masks the corresponding label data according to the new target class of interest in the target semantic space, that is, converts the new target class label to the background class, thereby establishing the source semantic space dataset; and the other part only retains the new target class label data of interest, and converts the rest to the background class, thereby establishing the target semantic space dataset;
[0010] Step S2: first, decouple the point cloud semantic segmentation model into a feature encoder and a classifier, and at the same time, construct a pre-trained semantic segmentation model and a target semantic segmentation model;
[0011] Step S3: Use the source semantic space point cloud information to perform supervised training on the pre-trained semantic segmentation model to obtain a reliable pre-trained semantic segmentation model for subsequent steps;
[0012] Step S4: Initialize the target semantic segmentation model using the pre-trained semantic segmentation model to achieve preliminary acquisition of the semantic segmentation capability of the target semantic segmentation model;
[0013] Step S5: Generate pseudo labels for the target semantic space point cloud based on the pre-trained semantic segmentation model, and use the pseudo labels to construct a target semantic space semantic segmentation training data set;
[0014] Step S6, freezing the parameters of the pre-trained semantic segmentation model, and using the pre-trained semantic segmentation model and the target semantic segmentation model to construct a distillation loss based on multi-level feature output;
[0015] Step S7, by limiting the feature space of the target semantic space point cloud obtained by the pre-trained semantic segmentation model and the target semantic segmentation model, and the generated pseudo-labels, the target semantic segmentation model is restricted to complete the training of the target semantic segmentation model.
[0016] Preferably, step S1 constructs a source point cloud semantic space dataset and a target point cloud semantic space dataset for the target semantic domain adaptive segmentation task by combining the existing vehicle-mounted laser point cloud semantic segmentation open source dataset and using the existing label data, specifically including the following steps:
[0017] Step A1: For a portion of the original open source dataset {P t-1 , L t-1}, retain the point cloud labels required for training the pre-trained point cloud semantic segmentation model, and attribute the point cloud labels that originally belong to the categories of interest in the target point cloud space to the background class labels, thereby forming a source point cloud semantic space dataset that does not contain new categories.
[0018] Step A2: Afterwards, for another part of the original open source dataset {P t , L t}, only the new class point clouds of interest in the target point cloud space are retained, and the remaining point cloud category labels are attributed to the background class labels, thereby forming a target point cloud semantic space dataset containing only new classes.
[0019] Preferably, step S2 specifically includes the following steps:
[0020] Step B1: Use a three-dimensional convolutional neural network (3DCNN) semantic segmentation model and decouple it into two parts, namely a feature encoder and a classifier.
[0021] Step B2: Construct a pre-trained semantic segmentation model, denoted as the t-1 time model, and its feature encoder is denoted as G t-1 , whose classifier is denoted by F t-1 After that, the target semantic space segmentation model is constructed, which is denoted as the t-time model, and its feature encoder is denoted as G t , whose classifier is denoted by F t .
[0022] Preferably, step S3 specifically includes the following steps:
[0023] Step C1: For a given source semantic space point cloud Using the feature encoder G of the original t-1 time model t-1 Calculate the point cloud P t-1 Feature representation f t-1 =G t-1 (P t-1 ), and regard it as a given point cloud P t-1 Feature space representation;
[0024] Step C2: Use the classifier F of the original t-1 model t-1 To classify the extracted last layer features, and use the cross entropy loss function L CE To constrain, so as to obtain a model with high confidence.
[0025] Preferably, step S4 specifically includes the following steps:
[0026] Step D1: According to the feature encoder G of the model at time t-1 t-1 The parameters are used to initialize the feature encoder G of the model at time t t, so that the feature encoder of the model at time t is consistent with that of the model at time t-1 at the initial stage.
[0027] Step D2: Classifier F based on the model at time t-1 t-1 The parameters of the old class are used to initialize the classifier F of the model at time t t The parameters of the old class make the model at time t consistent with the old class parameters of the model classifier at time t-1 at the initialization.
[0028] Step D3: Initialize the parameters of the new class in the classifier at time t according to the parameters of the background class in the classifier at time t-1, so that the classifier at time t has some ability to distinguish the new class at the beginning.
[0029] Preferably, step S5 specifically includes the following steps:
[0030] Step E1: Transform the target semantic space point cloud Input into the model at time t-1, and according to its output result, the result with high confidence is used as the pseudo label of the corresponding point cloud, denoted as L1={l t ,…,l N}.
[0031] Step E2: Utilize the new class point cloud in the target semantic space point cloud The true label L2 = {l n ,…,l N'} Update the pseudo label L1 generated at time t-1 = {l t ,…,l N}, making it a complete dataset L with new and old class labels t = {l t ,…,lN '}.
[0032] Step E3: Use the constructed complete dataset as the real data of the target semantic space point cloud as the model training set P at time t t , L t}.
[0033] Preferably, step S6 specifically includes the following steps:
[0034] Step F1: Freeze the feature encoder G of the model at time t-1 t-1 With classifier F t-1 And input the target semantic space point cloud into the model at time t-1 to obtain the feature representation f' t-1 ;
[0035] Step F2: Train the feature encoder G of the model at time t t With classifier F t , and input the target semantic space point cloud into the model at time t to obtain the feature representation ft , the highest confidence in the output result is taken as the pseudo label, and the label classified as the new class is attributed to the background class, denoted as L t-1 = {l' t ,…,l' N};
[0036] Step F3: Calculate the L2 norm of each level of feature output between the model at time t-1 and the model at time t, denoted as L Dis , the specific form is as follows:
[0037]
[0038] Where M is the number of layers output by the model feature encoder;
[0039] Preferably, step S7 specifically includes the following steps:
[0040] Step G1: Use the model output at time t-1 and L t-1 Label calculation cross entropy loss, denoted as L CE1 ;
[0041] Step G2: Use the model output at time t and L t Label calculation cross entropy loss, denoted as L CE2 ;
[0042] Step G3: Use L CE1 , L CE2 The loss cross coupling with L Dis The total loss L Overall , the specific form is as follows:
[0043] L Overall =L Dis +L CE1 +L CE2
[0044] Use L Overall To constrain the training of the model at time t, and finally expand the recognition capability of the original model.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] By decoupling the point cloud semantic segmentation model into a feature encoder and a classifier, a pre-trained semantic segmentation model is constructed based on the source semantic point cloud space; the target point cloud semantic segmentation model is initialized by the pre-trained point cloud semantic segmentation model to establish the target point cloud semantic segmentation model; the pre-trained point cloud semantic segmentation model is used to generate pseudo labels based on the target point cloud semantic space, and the target point cloud semantic space is constructed together with the real labels of the target point cloud semantic space; the feature representation of the target point cloud semantic space in the pre-trained point cloud semantic segmentation model and the target point cloud semantic segmentation model is constructed and restricted; a cross-coupling loss function is constructed based on the pre-trained point cloud semantic segmentation model and the target point cloud semantic segmentation model, so that the point cloud semantic segmentation model can learn new categories without forgetting the knowledge of old categories.
[0047] Theoretically, the method described in the present invention can solve the incremental learning problem of any form of point cloud semantic segmentation model, and can effectively enhance the universality of the incremental method of point cloud semantic segmentation. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 is a flow chart of the method of the present invention;
[0049] Figure 2 A schematic diagram of the multi-level feature space representation maintenance of the present invention;
[0050] Figure 3 This is a schematic diagram of the cross-coupling loss based on pseudo labels in the present invention. DETAILED DESCRIPTION
[0051] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.
[0052] like Figure 1-3 As shown, a vehicle-mounted laser point cloud target semantic domain adaptive semantic segmentation method can incrementally learn the point cloud semantic segmentation model using only a pre-trained point cloud semantic segmentation model and a target semantic space point cloud without relying on a source semantic point cloud, including the following steps;
[0053] Step S1. Therefore, this project first divides the original open source dataset into two parts by combining the existing vehicle-mounted laser point cloud semantic segmentation open source dataset. One part masks the corresponding label data according to the new target class of interest in the target semantic space, that is, converts the new target class label to the background class, thereby establishing the source semantic space dataset; and the other part only retains the new target class label data of interest, and converts the rest to the background class, thereby establishing the target semantic space dataset;
[0054] Step S2: first, decouple the point cloud semantic segmentation model into a feature encoder and a classifier, and at the same time, construct a pre-trained semantic segmentation model and a target semantic segmentation model;
[0055] Step S3: Use the source semantic space point cloud information to perform supervised training on the pre-trained semantic segmentation model to obtain a reliable pre-trained semantic segmentation model for subsequent steps;
[0056] Step S4: Initialize the target semantic segmentation model using the pre-trained semantic segmentation model to achieve preliminary acquisition of the semantic segmentation capability of the target semantic segmentation model;
[0057] Step S5: Generate pseudo labels for the target semantic space point cloud based on the pre-trained semantic segmentation model, and use the pseudo labels to construct a target semantic space semantic segmentation training data set;
[0058] Step S6, freezing the parameters of the pre-trained semantic segmentation model, and using the pre-trained semantic segmentation model and the target semantic segmentation model to construct a distillation loss based on multi-level feature output;
[0059] Step S7, by limiting the feature space of the target semantic space point cloud obtained by the pre-trained semantic segmentation model and the target semantic segmentation model, and the generated pseudo-labels, the target semantic segmentation model is restricted to complete the training of the target semantic segmentation model.
[0060] Step S1 combines the existing vehicle-mounted laser point cloud semantic segmentation open source dataset with the existing label data to construct a source point cloud semantic space dataset and a target point cloud semantic space dataset for the target semantic domain adaptive segmentation task, specifically including the following steps:
[0061] Step A1: For a portion of the original open source dataset {P t-1 , L t-1}, retain the point cloud labels required for training the pre-trained point cloud semantic segmentation model, and attribute the point cloud labels that originally belong to the categories of interest in the target point cloud space to the background class labels, thereby forming a source point cloud semantic space dataset that does not contain new categories.
[0062] Step A2: Afterwards, for another part of the original open source dataset {P t , L t}, only the new class point clouds of interest are retained in the target point cloud space, and the point cloud category labels originally belonging to the source semantic point cloud space are attributed to the background class labels, thereby forming a target point cloud semantic space dataset containing only new classes.
[0063] Step S2 specifically includes the following steps:
[0064] Step B1: Use a three-dimensional convolutional neural network (3DCNN) semantic segmentation model and decouple it into two parts, namely a feature encoder and a classifier.
[0065] Step B2: Construct a pre-trained semantic segmentation model, denoted as the t-1 time model, and its feature encoder is denoted as G t-1 , whose classifier is denoted by F t-1 After that, the target semantic space segmentation model is constructed, which is denoted as the t-time model, and its feature encoder is denoted as G t , whose classifier is denoted by F t .
[0066] Step S3 specifically includes the following steps:
[0067] Step C1: For a given source semantic space point cloud Using the feature encoder G of the pre-trained semantic segmentation model t-1 Calculate the point cloud P t-1 Feature representation f t-1 =G t-1 (P t-1 ), and regard it as a given point cloud P t-1 Feature space representation;
[0068] Step C2: Use the classifier F of the original t-1 model t-1 To classify the extracted last layer features, and use the cross entropy loss function L CE To constrain, so as to obtain a model with high confidence.
[0069] Step S4 specifically includes the following steps:
[0070] Step D1: According to the feature encoder G of the model at time t-1 t-1 The parameters are used to initialize the feature encoder G of the model at time t t , so that the feature encoder of the model at time t is consistent with that of the model at time t-1 at the initial stage.
[0071] Step D2: Classifier F based on the model at time t-1 t-1 The parameters are used to initialize the classifier F of the model at time t t The parameters of the old class make the model at time t consistent with the old class parameters of the model classifier at time t-1 at the initialization.
[0072] Step D3: Initialize the parameters of the new class in the classifier at time t according to the parameters of the background class in the classifier at time t-1, so that the classifier has some ability to distinguish the new class at the beginning.
[0073] Step S5 specifically includes the following steps:
[0074] Step E1: Transform the target semantic space point cloud Input into the model at time t-1, and according to its output result, the result with high confidence is used as the pseudo label of the corresponding point cloud, denoted as L1={l t ,…,l N}.
[0075] Step E2: Utilize the new class point cloud in the target semantic space point cloud The true label L2 = {l n ,…,l N'}Update the pseudo-label {l based on the generated pseudo-label at time t-1 t ,…,l N}, making it a complete dataset L with new and old class labels t = {l t ,…,l N'}.
[0076] Step E3: Use the constructed complete dataset as the real data of the target semantic space point cloud as the model training set P at time t t , L t}.
[0077] Step S6 specifically includes the following steps:
[0078] Step F1: Freeze the feature encoder G of the model at time t-1 t-1 With classifier F t-1 And input the target semantic space point cloud into the model at time t-1 to obtain the feature representation f' t-1 ;
[0079] Step F2: Train the feature encoder G of the model at time t t With classifier F t , and input the target semantic space point cloud into the model at time t to obtain the feature representation f t , the highest confidence in the output result is taken as the pseudo label, and the label classified as the new class is attributed to the background class, denoted as L t-1 = {l' t ,…,l' N}; This expression generates all pseudo labels containing the new class and replaces the pseudo labels belonging to the new class with the labels of the background class; that is, labels for all categories except the new class are generated.
[0080] Step F3: Calculate the L2 norm of each level of feature output between the model at time t-1 and the model at time t, denoted as L Dis , the specific form is as follows:
[0081]
[0082] Where M is the number of layers output by the model feature encoder;
[0083] Step S7 specifically includes the following steps:
[0084] Step G1: Use the model output at time t-1 and L t-1 Label calculation cross entropy loss, denoted as L CE1 ;
[0085] Step G2: Use the model output at time t and L t Label calculation cross entropy loss, denoted as L CE2 ;
[0086] Step G3: Use L CE1 , L CE2 The loss cross coupling with L Dis The total loss L Overall , the specific form is as follows:
[0087] L Overall =L Dis +L CE1 +L CE2
[0088] Use L Overall To constrain the training of the model at time t, and finally expand the recognition capability of the original model.
[0089] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "include..." or "comprise..." do not exclude the existence of other elements in the process, method, article or terminal device including the elements. In addition, in this article, "greater than", "less than", "exceed" and the like are understood to exclude the number itself; "above", "below", "within" and the like are understood to include the number itself.
[0090] Those skilled in the art should understand that the above embodiments can be provided as methods, devices, or computer program products. These embodiments can be in the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware. All or part of the steps in the methods involved in the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a storage medium readable by a computer device for executing all or part of the steps described in the above embodiments. The computer device includes, but is not limited to: personal computers, servers, general-purpose computers, special-purpose computers, network devices, embedded devices, programmable devices, smart mobile terminals, smart home devices, wearable smart devices, in-vehicle smart devices, etc.; the storage medium includes, but is not limited to: RAM, ROM, disk, tape, CD, flash memory, U disk, mobile hard disk, memory card, memory stick, network server storage, network cloud storage, etc.
[0091] The above embodiments are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a computer device to generate a machine, so that the instructions executed by the processor of the computer device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0092] These computer program instructions may also be stored in a computer device readable memory capable of directing a computer device to operate in a specific manner, so that the instructions stored in the computer device readable memory produce a manufactured product including an instruction device, the instruction device implementing the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0093] These computer program instructions can also be loaded onto a computer device so that a series of operating steps are executed on the computer device to produce a computer-implemented process, whereby the instructions executed on the computer device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0094] Although the above embodiments have been described, once those skilled in the art know the basic creative concepts, they can make additional changes and modifications to these embodiments. Therefore, the above description is only an embodiment of the present invention and does not limit the patent protection scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the specification and drawings of the present invention, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A semantic domain adaptive semantic segmentation method for vehicle-mounted laser point cloud targets, characterized by: The steps include: Step S1, combining the existing vehicle-mounted laser point cloud semantic segmentation open source dataset, the original open source dataset is divided into two parts as a whole, one of which masks the corresponding label data according to the new target class of interest in the target semantic space, that is, converts the new target class label to the background class, thereby establishing the source semantic space dataset; and the other part only retains the new target class label data of interest, and converts all the rest to the background class, thereby establishing the target semantic space dataset; Step S2: first, decouple the point cloud semantic segmentation model into a feature encoder and a classifier, and at the same time, construct a pre-trained semantic segmentation model and a target semantic segmentation model; Step S3: Use the source semantic space point cloud information to perform supervised training on the pre-trained semantic segmentation model to obtain a reliable pre-trained semantic segmentation model for subsequent steps; Step S4: Initialize the target semantic segmentation model using the pre-trained semantic segmentation model to achieve preliminary acquisition of the semantic segmentation capability of the target semantic segmentation model; Step S5: Generate pseudo labels for the target semantic space point cloud based on the pre-trained semantic segmentation model, and use the pseudo labels to construct a target semantic space semantic segmentation training data set; Step S6, freezing the parameters of the pre-trained semantic segmentation model, and using the pre-trained semantic segmentation model and the target semantic segmentation model to construct a knowledge distillation method based on multi-level feature preservation; Step S7, restricting the feature space of the target semantic space point cloud obtained by restricting the pre-trained semantic segmentation model and the target semantic segmentation model, and the generated pseudo-label to restrict the target semantic segmentation model, thereby completing the training of the target semantic segmentation model.
2. The method for adaptive semantic segmentation of vehicle-mounted laser point cloud target semantic domain according to claim 1, characterized in that: Step S1 combines the existing vehicle-mounted laser point cloud semantic segmentation open source dataset with the existing label data to construct a source semantic space dataset and a target semantic space dataset for the target semantic domain adaptive segmentation task; specific The steps include: Step A1: For a part of the original open source dataset {P t-1 , L t-1 }, retain the point cloud labels required for training the pre-trained semantic segmentation model, and classify the point cloud labels that originally belong to the category of interest in the target point cloud space into the background class labels, thereby forming a source semantic space dataset that does not contain new classes; Step A2: For another part of the original open source dataset {P t , L t }, only the new class point cloud labels of interest in the target point cloud space are retained, and the remaining point cloud category labels are classified as background category labels, thereby forming a target semantic space dataset containing only new classes.
3. The method for semantic domain adaptive semantic segmentation of vehicle-mounted laser point cloud targets according to claim 2, characterized in that: Step S2 specifically The steps include: Step B1: Use a 3D convolutional neural network (3DCNN) semantic segmentation model and decouple it into two parts, namely, a feature encoder and a classifier; Step B2: Construct a pre-trained semantic segmentation model, denoted as the t-1 time model, and its feature encoder is denoted as G t-1 , whose classifier is denoted by F t-1 ; Construct the target semantic space segmentation model, denoted as the t-time model, and its feature encoder is denoted as G t , whose classifier is denoted by F t .
4. The method for adaptive semantic segmentation of vehicle-mounted laser point cloud target semantic domain according to claim 3, characterized in that: Step S3 specifically The steps include: Step C1: For a given source semantic space point cloud Using the feature encoder G of the original t-1 time model t-1 Calculate the point cloud P t-1 Feature representation f t-1 =G t-1 (P t-1 ), and regard it as a given point cloud P t-1 Feature space representation; Step C2: Use the classifier F of the original t-1 model t-1 To classify the extracted last layer features, and use the cross entropy loss function L CE To constrain, so as to obtain a model with high confidence.
5. The method for adaptive semantic segmentation of vehicle-mounted laser point cloud target semantic domain according to claim 4, characterized in that: Step S4 specifically includes the following steps: Step D1: According to the feature encoder G of the model at time t-1 t-1 The parameters are used to initialize the feature encoder G of the model at time t t , so that the feature encoder of the model at time t is consistent with that of the model at time t-1 at the initial stage; Step D2: Classifier F based on the model at time t-1 t-1 The old class parameters are used to initialize the classifier F of the model at time t t The parameters of the old class make the model at time t consistent with the old class parameters of the model classifier at time t-1 at the initialization; Step D3: Initialize the parameters of the new class in the classifier at time t according to the parameters of the background class in the classifier at time t-1, so that the classifier at time t has some ability to distinguish the new class at the beginning.
6. The method for semantic domain adaptive semantic segmentation of vehicle-mounted laser point cloud objects according to claim 5, characterized in that: Step S5 specifically includes the following steps: Step E1: Transform the target semantic space point cloud Input into the model at time t-1, and according to its output result, the result with high confidence is used as the pseudo label of the corresponding point cloud, denoted as L1={l t ,…,l N }; Step E2: Utilize the new class point cloud in the target semantic space point cloud The true label L2 = {l n ,…,l N' } Update the pseudo label L1 generated at time t-1 = {l t ,…,l N }, making it a complete dataset L with new and old class labels t = {l t ,…,l N' }; Step E3: Use the constructed complete dataset as the real data of the target semantic space point cloud and as the model training set {P t , L t }.
7. The method for semantic domain adaptive semantic segmentation of vehicle-mounted laser point cloud targets according to claim 6, characterized in that: Step S6 specifically includes the following steps: Step F1: Freeze the feature encoder G of the model at time t-1 t-1 With classifier F t-1 And input the target semantic space point cloud into the model at time t-1 to obtain the feature representation f' t-1 ; Step F2: Train the feature encoder G of the model at time t t With classifier F t , and input the target semantic space point cloud into the model at time t to obtain the feature representation f t , the highest confidence in the output result is taken as the pseudo label, and the label classified as the new class is attributed to the background class, denoted as L t-1 = {l' t ,…,l' N }; Step F3: Calculate the L2 norm of each level of feature output between the model at time t-1 and the model at time t, denoted as L Dis , the specific form is as follows: Among them, M is the number of layers of the model feature encoder output.
8. The method for semantic domain adaptive semantic segmentation of vehicle-mounted laser point cloud targets according to claim 7, characterized in that: Step S7 specifically The steps include: Step G1: Use the model output at time t-1 and L t-1 Label calculation cross entropy loss, denoted as L CE1 ; Step G2: Use the model output at time t and L t Label calculation cross entropy loss, denoted as L CE2 ; Step G3: Use L CE1 , L CE2 The loss cross coupling with L Dis The total loss L Overall , the specific form is as follows: L Overall =L Dis +L CE1 +L CE2 Use L Overall To constrain the training of the model at time t, and finally expand the recognition capability of the original model.