Automatic driving system small sample increment target recognition learning method in long-tail environment

Abstract

The invention discloses an automatic driving system small sample increment target recognition learning method in a long-tail environment. The method comprises the following steps: training a target detection model on sufficient old category data as a teacher model; performing data playback enhancement by adopting a small sample playback method, and generating a composite image containing new and old category targets; adopting a Puzzlemix data enhancement method to generate a synthetic sample containing new and old category objects; the sample diversity is increased by adopting an Albumentations data enhancement method; constructing a feature enhancement module and introducing an attention mechanism to enhance old category feature distribution; training a student model by adopting double-stage distillation; and carrying out model convergence judgment and deployment updating. According to the method, rapid learning of a new target category in an automatic driving scene can be realized, the old category recognition precision is effectively maintained, and rapid deployment and overall recognition performance can be ensured when few sample target categories are newly added.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and machine learning technology, and relates to a small-sample incremental target recognition learning method for an autonomous driving system in a long-tail environment. Background Art

[0002] Current object detection models in autonomous driving systems rely heavily on large amounts of labeled data for offline training. The introduction of new categories often requires retraining the entire model by mixing new and old data, which presents numerous limitations in real-world scenarios. This is particularly true in applications subject to long-tail data distribution, where a large number of categories have only a handful of labeled examples. This severely restricts the scalability and practicality of object detection models. Few-shot learning (FSL) enables models to rapidly adapt to new tasks or recognize new categories with only a small number of examples. Incremental learning (IL) focuses on enabling models to continuously learn new knowledge while retaining old knowledge, effectively avoiding the problem of "catastrophic forgetting." Combining few-shot and incremental learning promises to enable models to continuously learn new categories with a small number of examples, while avoiding retraining on the full data set. This improves their flexibility and efficiency in real-world applications. However, few-shot incremental learning still has numerous limitations. Effectively mitigating forgetting in the face of imbalanced data and shifting distributions remains a key challenge in this field. Summary of the Invention

[0003] The purpose of the present invention is to provide a small-sample incremental target recognition learning method for an autonomous driving system in a long-tail environment, which can realize the rapid learning of new target categories in autonomous driving scenarios and effectively maintain the recognition accuracy of old categories. It can be quickly deployed and guarantee the overall recognition performance when adding new small-sample target categories.

[0004] The technical solution adopted in the present invention is: The small sample incremental target recognition learning method for autonomous driving systems in long-tail environments is implemented in the following steps: Step 1: Train the target detection model as a teacher model on sufficient old category data; Step 2: Use a small sample playback method to perform data playback enhancement to generate a composite image containing both new and old category targets; Step 3: Use the Puzzlemix data augmentation method to generate synthetic samples containing objects of new and old categories; Step 4: Use Albumentations data enhancement method to increase sample diversity; Step 5: Construct a feature enhancement module and introduce an attention mechanism; Step 6: Using the feature enhancement module constructed in step 5, the student model is trained using two-stage distillation. Step 7: Model convergence determination and deployment update.

[0005] The beneficial effects of the present invention are: (1) The method of the present invention adopts a small sample incremental learning mechanism, which enables the target recognition model to quickly adapt to and recognize new target categories without retraining the full amount of data, thereby improving the adaptability and scalability of the model; (2) The proposed method constructs a teacher-student dual network architecture and introduces a two-stage distillation strategy (feature distillation and logic distillation) to effectively transfer old category knowledge, significantly alleviate the catastrophic forgetting problem, and ensure that the model maintains the recognition accuracy of old categories while learning new categories; (3) The method of the present invention adopts a dual enhancement strategy (data enhancement and feature enhancement), namely, introducing a variety of image enhancement methods such as PuzzleMix and Albumentations, to improve the generalization ability of the model to complex environments such as occlusion, rain and fog, and lighting changes, thereby enhancing the practicality and robustness of the model; (4) The proposed method combines the Copy-Paste synthetic sample with the feature enhancement strategy to effectively enhance the distinction between new and old categories in the image and feature space, and improve the recognition accuracy of the model for similar targets; (5) The proposed method only requires a small number of new samples to complete the model update, avoiding the high computational and storage costs of reloading and training all the data. It has good real-time performance and engineering application value, and solves practical engineering problems faced by autonomous driving systems in target recognition in long-tail road environments, such as sample scarcity, low recognition accuracy, difficulty in online updates, and limited on-board computing power. Experiments on datasets such as BDD100K have shown that the proposed method, using only a small number of new class samples, can improve the mAP@50 of new classes to 50.72%, and the accuracy drop of old classes is controlled within 5%, significantly outperforming traditional fine-tuning and existing small-sample incremental methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 It is an overall flow chart of the method of the present invention; Figure 2 This is a dual-enhanced small sample incremental learning DA-FSIL model diagram based on a teacher-student dual network architecture proposed by the method of the present invention; Figure 3 This is an example diagram of the PuzzleMix data enhancement method of the present invention; Figure 4 This is an example diagram of the Albumentations data enhancement method of the method of the present invention; Figure 5It is part of the original data of the new and old class objects used in training the model in the incremental learning of the new class "bus" on urban roads in Example 1 of the present invention. DETAILED DESCRIPTION

[0007] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0008] The present invention provides a small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment, such as Figure 1 and Figure 2 As shown, please follow the steps below: Step 1: Teacher model construction: The YOLOv8s object detection model is fully trained on the old category image dataset until stable and high-precision detection performance is achieved on the validation set, and this model is used as the teacher model.

[0009] Step 2: Use small sample playback method to enhance data playback: 3-5 representative objects (with high image confidence, such as 0.9) are selected from the old class dataset and stored as the old class replay data. During the data replay process, a cropped object is randomly selected from the memory bank and its width and height are randomly adjusted within a certain range. Then, an object of the new class is selected from the new sample and the old and new objects are pasted onto a clean background using the copy-paste technique. During this process, the intersection over union (IoU) between the new and old objects is kept below 0.3 to ensure that the overlap between the new and old objects is not too large, thus avoiding occlusion. Ultimately, a composite image containing objects of the new and old categories is generated.

[0010] Step 3: Generate synthetic samples using the Puzzlemix data augmentation method: The new and old class images are used as input, and the PuzzleMix data enhancement method is used to generate synthetic samples containing new and old category objects, thereby improving the model's ability to perceive potential target objects. Figure 3 As shown in Figure 1, Figure a is the old category data, Figure b is the new category data, and Figure c is a synthetic sample containing new and old category objects generated by the PuzzleMix method.

[0011] The steps of the PuzzleMix method are: (1) Region selection and mask generation: Generate a saliency map using a saliency detection method (gradient-based saliency map) , respectively identify the input image and Then, based on these salient regions, an irregularly shaped binary mask is generated for each image, where the salient regions correspond to a mask value of 1 (retained) and the background regions correspond to 0 (possibly replaced).

[0012] Saliency Map The generation is shown in formula (1): (1) Where: ——Represents the loss function, which measures the model’s input x The difference between the output and the target; ——Indicates the input image x gradient.

[0013] (2) Region alignment based on optimal transmission: Calculate image The salient areas and images The characteristic distance between the salient regions is calculated by using the optimal transmission theory (Sinkhorn algorithm) to solve a transmission plan that minimizes the overall matching cost. The significant area is "transmitted" to The optimal alignment at the region level is achieved by aligning the salient regions of the image (or vice versa). This alignment process determines Which areas should correspond to The formula to minimize the overall matching cost is as follows: (2) Where: and ——represents two input images; and ——Indicates that the significant area will be moved from and The transmission plan matrix for transmission into the mixed sample; ——Represents the transmission cost matrix, which measures the cost of data movement; ——Represents the weight coefficient, which controls the impact of transmission cost in optimization; ——Represents a matrix and the cost matrix The inner product of ——Indicates the optimal mixing mask; ——Indicates the input image and The salient map.

[0014] (3) Hybrid image generation: Adjust the hybrid mask according to the optimal alignment scheme, and use the optimized hybrid mask to perform a weighted combination of the salient areas of the two input images, ultimately generating a hybrid sample (hybrid image) that retains key object information and promotes the model to learn discriminative features. The generation is shown in formula (3): (3) Where: ——Represents the mixed mask vector is a n dimensional vector, each element value is a real number (continuous value) between 0 and 1, used to control the mixing ratio of each area; ——Represents the transfer plan matrix, which is used to adjust the matching of salient regions.

[0015] In the mixed sample construction, PuzzleMix optimizes the mixed mask To balance the exposure of the salient areas, thereby maximizing the appearance of salient information while maintaining the structural integrity of the data and the local smoothness of the image. This process can be formalized as an optimization problem of the mixed mask, the goal is to minimize the following energy function: (4) Where: ——Represents the mixed mask vector belong N dimensional state space ; ——Represents the mixed mask vector No. , Component, that is, , The status of each location; ——Indicates the regional significance item, which is used to measure the selection of each region The significance of the post- ——Represents the label smoothing term, which is used to control the smoothness of the mixing ratio; ——Represents the data smoothing term, which is used to control the smoothness of adjacent areas; ——represents the prior term, which is used to control the prior distribution of the mixing ratio; ——represents a set of adjacent areas; ——Represents the weight coefficient, which is used to adjust the influence of each item.

[0016] (4) Loss calculation: The generated mixed image is input into the model for forward propagation and the loss function is calculated. The loss is calculated by weighting the original image label according to the ratio of the mixed mask.

[0017] Step 4: Use Albumentations data enhancement method to increase sample diversity: The Albumentations method is used to perform efficient and diverse data enhancement on the original old and new class images, the small sample playback processed images obtained in step 2, and the images generated by the PuzzleMix method in step 3. A variety of transformation methods, such as rotation, illumination change, and noise adjustment, are used to generate diverse training samples, thereby enhancing the model's adaptability to different scenarios. The CoarseDropout method is used to simulate different degrees of occlusion to enhance the model's ability to recognize occluded targets. At the same time, methods such as RandomRain and RandomFog are used to simulate rainy and foggy weather, thereby improving the robustness of the model in various scenarios. Figure 4 As shown, Figure a is the original image, and Figures b and d are enhanced with CoarseDropout occlusion, rainy day, and foggy scene simulation respectively.

[0018] Step 5: Build a feature enhancement module and introduce the attention mechanism: Construct a feature enhancement module and use the distribution information of the old category (category mean and covariance ) to enhance the feature space and improve the representation of old category features by regularizing the classifier learning process. At the same time, the representation of new categories is used to enrich the distribution of variants of old categories to distinguish old and new categories with similar appearance, such as "bus" and "truck" and "bicycle" and "motorcycle".

[0019] The principle of feature enhancement method is: First, use the old model to build a relationship graph between the old and new categories ,in is a category node, is the similarity between the old category and the new category. Similarity Indicates a new category Predicted as old category The ratio is calculated by the following formula: (5) Where: ——Indicates the indicator function; ——Indicates the predicted category of the sample by the old model; ——Indicates the number of samples of the new category; ——Category label indicating the old category.

[0020] Then, the confusion matrix is ​​calculated using the new category data to obtain the covariance information of the old category, and the feature information of the new category is combined to improve the distribution of the old category. The covariance after fusion is calculated using the following formula: (6) Where: ——Indicates a new category The covariance matrix of ——Indicates the old category The covariance matrix of .

[0021] In order to further enhance feature expression, the attention mechanism is introduced. The attention mechanism can adaptively focus on the key areas and channels in the feature map, thereby optimizing the feature enhancement process. Channel attention and spatial attention are used in the feature enhancement module to emphasize the important channels and spatial areas in the feature map respectively. First, the channel attention weight is calculated by global average pooling, convolution operation and Sigmoid function. , used to control the importance of each channel, such as the formula: (7) Where: ——Indicates the Sigmoid activation function; ——represents the input feature map; ——Indicates global average pooling; ——Indicates the convolution operation.

[0022] Then the spatial attention weight is calculated by fusion information through average pooling and maximum pooling , thereby controlling the weights of important areas in the feature map: (8) Where: ——Indicates that the average pooling operation is performed on the input feature map; ——Indicates the maximum pooling operation on the input feature map; ——Indicates splicing operation.

[0023] After combining channel and spatial attention, we get enhanced feature representation : (9) Afterwards, the feature representation after attention enhancement is Perform Gaussian sampling to obtain the final enhanced features : (10) Where: ——Indicates the number of channels, that is, the depth dimension of the input feature map.

[0024] The feature enhancement loss is defined as follows: (11) Where: and ——Indicates the k Class weights and biases; ——Indicates the number of enhanced features generated; and ——represent the number of old categories and the number of all categories respectively.

[0025] Step 6: Use two-stage distillation to train the student model: The student model is trained iteratively using samples processed in steps 2, 3, and 4. To further preserve the learned feature representations of the old categories while maintaining discrimination across all categories, a two-stage distillation process is introduced to compensate for the deficiencies in the old class representations in the feature space. The first stage is feature distillation, in which the student model learns representations by comparing them with the teacher model's feature maps. This is then combined with the feature enhancement module constructed in step 5 to enrich the feature representations of the old categories. The second stage is logistic distillation, in which the student model aligns with the teacher model's soft labels using the Kullback-Leibler (KL) divergence loss, learning its output distribution and thus preserving knowledge of the old categories.

[0026] Training loss during feature distillation Mainly due to feature distillation loss and feature enhancement loss and standard detection loss The composition is as follows: (12) Where: ——represents the weight of feature distillation loss; ——represents the weight of feature enhancement loss.

[0027] Training loss during the logistic distillation phase Mainly by standard detection loss and logistic distillation loss The composition is as follows: (13) Where: ——Represents the weight of the logistic distillation loss.

[0028] Step 7: Model convergence determination and deployment update: During training, the student model's mAP@50 and old-class accuracy are continuously monitored on a validation set containing both new and old categories. The model is considered converged when the new-class mAP@50 reaches a set threshold (e.g., above 50%) and the old-class detection accuracy decreases to an acceptable level (e.g., no more than 5%). At this point, the trained student model replaces the original teacher model and is deployed in the autonomous driving perception system, retaining its incremental learning capability. By repeating the seven steps above and continuously introducing new categories, the model can be continuously updated online and its capabilities expanded.

[0029] Example 1: This example demonstrates incremental learning of a new category of “buses” on urban roads. The specific steps are as follows: Step 1: Teacher model construction: In this embodiment, a training dataset containing 6 old categories was constructed based on the BDD100K dataset, including pedestrians, cars, trucks, bicycles, traffic lights, and traffic signs. A total of 8,000 valid images were screened out and divided into a training set (5,600 images), a validation set (1,200 images), and a test set (1,200 images) in a ratio of 7:1.5:1.5. At the same time, buses were extracted from BDD100K as a new category, and 20 instances were selected to form a small sample new dataset of 20 images. The same ratio of 7:1.5:1.5 was used to divide the original data of the new and old class objects during training. Figure 5 As shown in Figure 1, Figure a shows the old class object, and Figure b shows the new class object. The teacher model is trained on the old class data based on YOLOv8s for 200 rounds, with an initial learning rate of 0.001 and a batch size of 12.

[0030] Step 2: Use small sample playback method to enhance data playback: In the old class dataset, we select 3-5 representative objects (with image confidence scores of at least 0.9) from each old class and store them as the old class replay data. We randomly select old class samples from the replay pool and fuse them with new class "bus" samples in a 1:1 ratio using Copy-Paste to a clean background, maintaining an Intersection over Union (IoU) between 0.2 and 0.3 to generate a composite image.

[0031] Step 3: Generate synthetic samples using the Puzzlemix data augmentation method: Taking new class (bus) and old class images (pedestrians, cars, trucks, bicycles, traffic lights and traffic signs) as input, the PuzzleMix data augmentation method is used to generate synthetic samples containing objects of new and old categories.

[0032] Step 4: Use Albumentations data enhancement method to increase sample diversity: Albumentations is used to perform efficient and diverse data augmentation on the original old and new class images, images processed by small sample playback, and images generated by the PuzzleMix method. For example, the CoarseDropout method is used to simulate different degrees of occlusion, while RandomRain and RandomFog methods are used to simulate rainy and foggy weather.

[0033] Step 5: Build a feature enhancement module and introduce the attention mechanism: A confusion matrix is ​​constructed based on the 20% probability that the teacher model misclassifies a bus as a "truck." This matrix is ​​then used to weight the covariance of the new class and integrate it into the feature distribution of the old class. A 1×1 channel attention convolution kernel and a 7×7 spatial attention convolution kernel are used to highlight key areas of the bus and the old class. Finally, Gaussian sampling is used to generate 10 enhanced feature vectors, which are then input into the student model.

[0034] Step 6: Use two-stage distillation to train the student model: The student model is trained iteratively using augmented samples from both the old and new classes. During training, the total loss is composed of detection loss, feature distillation loss, and logistic distillation loss. Training is conducted in two phases, totaling 80 rounds: the first 40 rounds are dedicated to feature distillation only, and the last 40 rounds incorporate logistic distillation.

[0035] Step 7: Model convergence determination and deployment update: After several training iterations, the student model's bus mAP@50 improved to 50.72% on the validation set containing both old and new classes, and the detection accuracy of the old class decreased by only 2.8%, demonstrating that the student model has good results in the incremental recognition of new classes on urban roads and can be used in place of the teacher model.

[0036] Example 2: The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment of this embodiment is specifically implemented according to the following steps: Step 1: Train the target detection model as a teacher model on sufficient old category data; Step 2: Use a small sample playback method to perform data playback enhancement to generate a composite image containing both new and old category targets; Step 3: Use the Puzzlemix data augmentation method to generate synthetic samples containing objects of new and old categories; Step 4: Use Albumentations data enhancement method to increase sample diversity; Step 5: Construct a feature enhancement module and introduce an attention mechanism; Step 6: Using the feature enhancement module constructed in step 5, the student model is trained using two-stage distillation. Step 7: Model convergence determination and deployment update.

[0037] Example 3: Based on Example 2, step 1 is specifically as follows: The object detection model is fully trained on the old category image dataset until stable and high-precision detection performance is achieved on the validation set, and this model is used as the teacher model.

[0038] Example 4: Based on Example 3, step 2 is specifically as follows: A representative small sample library is formed by screening 3 to 5 samples of each category with a confidence level of not less than 0.9 from the old category dataset. By pasting the old and new category targets into a unified clean background, a composite image containing the old and new category targets is generated to enhance category discrimination, where the intersection-over-union ratio of the new and old targets is less than 0.3.

[0039] Example 5: Based on Example 4, step 3 is specifically as follows: The new and old class images are used as input to PuzzleMix. First, a saliency map is generated using a saliency detection method. , respectively identify the input image and The most informative salient regions in the image are identified. Based on these salient regions, an irregularly shaped binary mask is generated for each image. The mask value of the salient region is 1, and the background region is 0. Saliency Map The generation is shown in formula (1): (1) Where: ——Represents the loss function, which measures the model’s inputx The difference between the output and the target; ——Indicates the input image x gradient; Secondly, calculate the image The salient areas and images The characteristic distance between the salient regions is calculated by using the optimal transmission theory to solve a transmission plan that minimizes the overall matching cost. Significant regional transmission to The optimal alignment at the regional level is achieved at the position of the significant area; the formula for minimizing the overall matching cost is as follows: (2) Where: and ——Indicates that the significant area will be moved from and The transmission plan matrix for transmission into the mixed sample; ——Represents the transmission cost matrix, which measures the cost of data movement; ——Represents the weight coefficient, which controls the impact of transmission cost in optimization; ——Represents a matrix and the cost matrix The inner product of ——Indicates the optimal mixing mask; ——Indicates the input image and salient map of Finally, the hybrid mask is adjusted according to the optimal alignment scheme, and the optimized hybrid mask is used to perform a weighted combination of the salient areas of the two input images to generate a hybrid image that not only retains key object information but also promotes the model to learn discriminative features; the hybrid image The generated formula is as follows: (3) Where: ——Represents the mixed mask vector is a n dimensional vector, each element value is a real number between 0 and 1, used to control the mixing ratio of each area; ——Represents the transfer plan matrix, which is used to adjust the matching of salient regions.

[0040] Example 6: Based on Example 5, in the mixed sample construction, PuzzleMix optimizes the mixed mask To balance the exposure of the salient area, thereby maximizing the appearance of salient information while maintaining the structural integrity of the data and the local smoothness of the image; this process is formalized as an optimization problem of the mixed mask, the goal is to minimize the following energy function: (4) Where: ——Represents the mixed mask vector belong N dimensional state space ; ——Represents the mixed mask vector No. , Component, that is, , The status of each location; ——Indicates the regional significance item, which is used to measure the selection of each region The significance of the post- ——Represents the label smoothing term, which is used to control the smoothness of the mixing ratio; ——Represents the data smoothing term, which is used to control the smoothness of adjacent areas; ——represents the prior term, which is used to control the prior distribution of the mixing ratio; ——represents a set of adjacent areas; ——Represents the weight coefficient, which is used to adjust the influence of each item.

Claims

1. A small sample incremental target recognition learning method for autonomous driving systems in long-tail environments, characterized by: Please follow the steps below to implement it: Step 1: Train the target detection model as a teacher model on sufficient old category data; Step 2: Use a small sample playback method to perform data playback enhancement to generate a composite image containing both new and old category targets; Step 3: Use the Puzzlemix data augmentation method to generate synthetic samples containing objects of new and old categories; Step 4: Use Albumentations data enhancement method to increase sample diversity; Step 5: Construct a feature enhancement module and introduce an attention mechanism; Step 6: Using the feature enhancement module constructed in step 5, the student model is trained using two-stage distillation. Step 7: Model convergence determination and deployment update.

2. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 1 is characterized in that: Step 1 is as follows: The object detection model is fully trained on the old category image dataset until stable and high-precision detection performance is achieved on the validation set, and this model is used as the teacher model.

3. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 1, characterized in that: Step 2 is as follows: A representative small sample library is formed by screening 3 to 5 samples of each category with a confidence level of not less than 0.9 from the old category dataset. By pasting the old and new category targets into a unified clean background, a composite image containing the old and new category targets is generated to enhance category discrimination, where the intersection-over-union ratio of the new and old targets is less than 0.

3.

4. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 1, characterized in that: Step 3 is as follows: The new and old class images are used as input to PuzzleMix. First, a saliency map is generated using a saliency detection method. , respectively identify the input image and The most informative salient regions in the image are identified. Based on these salient regions, an irregularly shaped binary mask is generated for each image. The mask value of the salient region is 1, and the background region is 0. Saliency Map The generated formula is as follows: Where: ——Represents the loss function, which measures the model’s input x The difference between the output and the target; ——Indicates the input image x gradient; Secondly, calculate the image The salient areas and images The characteristic distance between the salient regions is calculated by using the optimal transmission theory to solve a transmission plan that minimizes the overall matching cost. Significant regional transmission to The optimal alignment at the regional level is achieved at the position of the significant area; the formula for minimizing the overall matching cost is as follows: Where: and ——Indicates that the significant area will be moved from and The transmission plan matrix for transmission into the mixed sample; ——Represents the transmission cost matrix, which measures the cost of data movement; ——Represents the weight coefficient, which controls the impact of transmission cost in optimization; ——Represents a matrix and the cost matrix The inner product of ——Indicates the optimal mixing mask; ——Indicates the input image and salient map of Finally, the hybrid mask is adjusted according to the optimal alignment scheme, and the optimized hybrid mask is used to perform a weighted combination of the salient areas of the two input images to generate a hybrid image that not only retains key object information but also promotes the model to learn discriminative features; the hybrid image The generated formula is as follows: Where: ——Represents the mixed mask vector is a n dimensional vector, each element value is a real number between 0 and 1, used to control the mixing ratio of each area; ——Represents the transfer plan matrix, which is used to adjust the matching of salient regions.

5. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 4 is characterized in that: In the mixed sample construction, PuzzleMix optimizes the mixed mask To balance the exposure of the salient area, thereby maximizing the appearance of salient information while maintaining the structural integrity of the data and the local smoothness of the image; this process is formalized as an optimization problem of the mixed mask, the goal is to minimize the following energy function: Where: ——Represents the mixed mask vector belong N dimensional state space ; ——Represents the mixed mask vector No. , Component, that is, , The status of each location; ——Indicates the regional significance item, which is used to measure the selection of each region The significance of the post- ——Represents the label smoothing term, which is used to control the smoothness of the mixing ratio; ——Represents the data smoothing term, which is used to control the smoothness of adjacent areas; ——represents the prior term, which is used to control the prior distribution of the mixing ratio; ——represents a set of adjacent areas; ——Represents the weight coefficient, which is used to adjust the influence of each item.

6. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 1, characterized in that: Step 4 is as follows: The Albumentations method is used to perform data enhancement on the original old and new class images, the small sample playback processed images obtained in step 2, and the images generated by the PuzzleMix method in step 3.

7. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 1, characterized in that: The workflow of the feature enhancement module constructed in step 5 is as follows: First, use the old model to build a relationship graph between the old and new categories ,in is a category node, is the similarity between the old category and the new category, similarity Indicates a new category Predicted as old category The ratio is calculated by the following formula: Where: ——Indicates the indicator function; ——Indicates the predicted category of the sample by the old model; ——Indicates the number of samples of the new category; ——category label indicating the old category; Secondly, the confusion matrix is ​​calculated using the new category data to obtain the covariance information of the old category, and the distribution of the old category is improved by combining the feature information of the new category. The covariance after fusion is calculated using the following formula: Where: ——Indicates a new category The covariance matrix of ——Indicates the old category The covariance matrix of Again, the channel and spatial attention mechanism is introduced. The channel attention weight is calculated by global average pooling, convolution operation and Sigmoid function. , used to control the importance of each channel, such as the formula: Where: ——Indicates the Sigmoid activation function; ——represents the input feature map; ——Indicates global average pooling; ——Indicates convolution operation; Afterwards, the spatial attention weight is calculated by fusion of information through average pooling and maximum pooling , thereby controlling the weights of important areas in the feature map: Where: ——Indicates that the average pooling operation is performed on the input feature map; ——Indicates the maximum pooling operation on the input feature map; ——Indicates splicing operation; After combining channel and spatial attention, we get enhanced feature representation : Finally, the feature representation after attention enhancement Perform Gaussian sampling to obtain the final enhanced features : Where: ——Indicates the number of channels, that is, the depth dimension of the input feature map.

8. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 7, characterized in that: The feature enhancement loss is defined as follows: Where: and ——Indicates the k Class weights and biases; ——Indicates the number of enhanced features generated; and ——represent the number of old categories and the number of all categories respectively.

9. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 1, characterized in that: In step 6, the student model is trained iteratively using the samples obtained after processing in steps 2, 3, and 4. A two-stage distillation process is introduced to compensate for the insufficient representation of the old categories in the feature space: the first stage is feature distillation. In this stage, the student model learns the representation by comparing with the feature map of the teacher model, and combines it with the feature enhancement module constructed in step 5 to enrich the feature expression of the old categories; the second stage is logical distillation. The student model aligns with the soft labels of the teacher model through the KL divergence loss and learns its output distribution, thereby retaining the knowledge of the old categories.

10. The small sample incremental target recognition learning method for an autonomous driving system in a long-tail environment according to claim 9, characterized in that: In step 6, the training loss of the feature distillation stage is Loss due to feature distillation Feature enhancement loss and standard detection loss The composition is as follows: Where: ——represents the weight of feature distillation loss; ——represents the weight of feature enhancement loss; Training loss during the logistic distillation phase Loss detected by standard and logistic distillation loss The composition is as follows: Where: ——Represents the weight of the logistic distillation loss.

Images