Traffic sign detection method based on single sample learning

Abstract

The invention discloses a traffic sign detection method based on single sample learning. The method comprises the following steps: constructing a training set, a verification set and a test set based on a traffic sign reference data set; for the training set and the verification set, adopting a pre-training-based open vocabulary target detection model to generate traffic sign candidate areas in the image, and carrying out intersection-to-union ratio screening on the traffic sign candidate areas to obtain traffic sign candidate areas related to the task and category labels; training the task-related traffic sign candidate areas and the category labels by adopting a single sample recognition algorithm based on metric learning to obtain a trained traffic sign detection model; processing the test image of the test set based on the trained traffic sign detection model to obtain a preliminary matching result; based on the preliminary matching result, a visual language model and semantic prompt guide words are adopted to obtain a semantic verification result, and based on the semantic verification result, a Kalman filtering and intersection-to-union matching strategy is adopted to obtain a final traffic sign detection result.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent transportation and computer vision, and in particular relates to a traffic sign detection method based on single-sample learning. Background Art

[0002] With the continuous development of intelligent transportation systems and autonomous driving technologies, traffic sign detection has become one of the key technologies. Traditional deep learning detection methods usually rely on a large amount of labeled data and show poor robustness in complex road environments (such as lighting changes, dynamic occlusion, and multi-target interference). In particular, when samples are scarce, the detection accuracy of rare signs drops significantly. It is worth noting that traffic signs have standardized visual features, and their templated characteristics have strong prior transferability. Therefore, how to use these standardized templates to achieve accurate positioning and recognition of traffic signs in images, break through the dependence on large-scale labeled data, and improve detection accuracy and robustness in complex scenarios has become a core issue that needs to be solved in the field of intelligent transportation.

[0003] Currently, mainstream traffic sign detection methods are mostly based on convolutional neural networks (CNNs), which rely heavily on large-scale annotated datasets during training. These methods struggle to maintain stable detection accuracy and robustness in small sample sizes or complex environments (such as changing lighting, object occlusion, and background interference). Meanwhile, with the continuous development of visual language models, open vocabulary object detection technology has been gradually applied to the field of object detection. This technology uses natural language cues to achieve target localization, but its positioning accuracy is low when processing fine-grained semantic cues (such as specific speed limit values), which has become a bottleneck for its application in traffic sign detection.

[0004] Therefore, the present invention proposes a traffic sign detection method based on single-sample learning, which aims to effectively solve the false detection problem of traditional methods in complex scenarios and reduce the dependence on large-scale labeled data. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a traffic sign detection method based on single-sample learning to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention provides a traffic sign detection method based on single-sample learning, comprising:

[0007] Construct training, validation, and test sets based on the traffic sign benchmark dataset;

[0008] Generate traffic sign candidate regions in the image using a pre-trained open vocabulary object detection model for the training set and the validation set, and perform intersection-over-union (IoU) screening on the traffic sign candidate regions to obtain task-related traffic sign candidate regions and category labels;

[0009] A single-sample recognition algorithm based on metric learning is used to train the candidate traffic sign regions and category labels related to the task to obtain a trained traffic sign detection model;

[0010] Processing the test image of the test set based on the trained traffic sign detection model to obtain a preliminary matching result;

[0011] Based on the preliminary matching results, the visual language model and semantic clue guide words are used to obtain the semantic verification results:

[0012] Based on the semantic verification result, a Kalman filter and intersection-over-union matching strategy is adopted to obtain the final traffic sign detection result.

[0013] Optionally, the process of constructing a training set, a validation set, and a test set based on the traffic sign benchmark dataset includes:

[0014] Eliminating sparse categories in the traffic sign benchmark dataset to construct a target category set;

[0015] Dividing the target class set into basic categories and novel categories, and constructing a category distribution system for small sample learning scenarios;

[0016] The training set, validation set, and test set are constructed based on the category distribution system in the small sample learning scenario.

[0017] Optionally, the process of obtaining task-related candidate traffic sign regions and category labels includes:

[0018] For the training set and the validation set, using a pre-trained open vocabulary object detection model and coarse-grained semantic cue words to obtain candidate traffic sign regions in the image;

[0019] Based on the intersection-over-union ratio between the traffic sign candidate region and the labeled box, the task-related traffic sign candidate region and category label are obtained.

[0020] Optionally, the trained traffic sign detection model is a single-sample traffic sign recognition network;

[0021] The framework of the single-sample traffic sign recognition network includes: a transferable feature encoder, a latent space encoder and a minimum distance classifier.

[0022] Optionally, the minimum distance classifier calculates the distance between the query image and the support image by a metric learning method to achieve classification and calculate the probability of each category;

[0023] Among them, the expression for calculating the probability of each category is:

[0024]

[0025] Among them, x is the feature vector of the input image, u k represents the prototype of the kth type of traffic sign, n represents the total number of prototypes of traffic signs, u j represents the prototype of the j-th type of traffic sign, and P(y=k|x) is the predicted probability that the image belongs to the k-th type of traffic sign.

[0026] Optionally, the process of obtaining a preliminary matching result based on the trained traffic sign detection model includes:

[0027] Based on the trained traffic sign detection model, the minimum distance between the area to be detected and the traffic sign template is obtained;

[0028] Based on the minimum distance, a preliminary matching result is obtained.

[0029] Optionally, the process of obtaining the final traffic sign detection result by adopting a Kalman filter and intersection-over-union matching strategy based on the semantic verification result includes:

[0030] constructing a target motion trajectory model through Kalman filtering, processing the preliminary matching result based on the target motion trajectory model, and obtaining a target motion trajectory prediction result;

[0031] Based on the target motion trajectory prediction result, an intersection-over-union matching strategy is adopted to obtain the category ID association result of the cross-frame candidate region;

[0032] The category ID association results based on the candidate regions across frames are processed using a majority voting mechanism to obtain the false detection correction result of the current frame;

[0033] The false detection correction results based on the current frame are integrated to obtain the final traffic sign detection result.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] This paper addresses the challenges of traditional convolutional neural networks (CNNs) in traffic sign detection, which rely heavily on large-scale annotated data and lack robustness in complex road scenarios. By introducing a single-shot learning-based traffic sign detection method, this method accurately locates and recognizes traffic signs in images by introducing category templates. Specifically, a pretrained open-vocabulary object detection model is used as input for coarse-grained semantic cues to generate candidate regions for traffic signs. Subsequently, a feature metric space is constructed based on a library of traffic sign templates. A template-driven matching algorithm is employed to calculate the minimum distance between the region to be detected and the template in the embedding space, achieving preliminary category matching for the candidate regions. Furthermore, a visual language model is incorporated into a semantic consistency verification module, which uses natural language guidance to perform semantic filtering on the preliminary matching results, effectively eliminating irrelevant objects. Finally, a temporal trajectory model is constructed using a Kalman filter algorithm. Cross-frame category consistency tracking and false detection correction for candidate regions are achieved using an intersection-over-union (IOU) correlation strategy, forming a closed-loop detection process consisting of "semantically guided candidate generation - template-driven feature matching - semantic verification and filtering - spatiotemporal trajectory correction." The present invention can complete target detection and classification of traffic signs by relying solely on category templates without relying on large-scale labeled samples, significantly improving the detection capability of rare categories and the robustness of the system in complex traffic environments. It is particularly suitable for application scenarios with high requirements on precision and stability, such as autonomous driving, road monitoring, and intelligent transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0037] Figure 1 Schematic diagram of novel categories selected based on the TT100K dataset under different grouping conditions according to an embodiment of the present invention;

[0038] Figure 2 This is a diagram of the overall framework of the network in an embodiment of the present invention;

[0039] Figure 3 Flowchart of the spatiotemporal trajectory correction module in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0042] Example 1

[0043] This paper addresses the current challenges of traffic sign detection, such as a strong reliance on large-scale annotated data, a high risk of false detection in complex scenarios, and insufficient ability to identify rare categories. By proposing a single-shot learning traffic sign detection method, this paper develops a pre-trained open-vocabulary object detection model and generates potential traffic sign candidate regions using coarse-grained semantic cues. This method then uses a metric learning-based single-shot template matching algorithm to rapidly classify candidate regions and determine the category of their nearest traffic sign template. A visual language model is then introduced to construct a semantic verification filtering module, which uses semantic cues to guide the model in determining the semantic match between candidate regions and templates, effectively filtering out task-irrelevant traffic signs. Finally, a target motion trajectory model is constructed based on a Kalman filter, and an intersection-over-union matching strategy is used to associate the category IDs of candidate regions across frames. Class information from historical frames is then used to correct false detections in the current frame. By focusing on four key steps: semantically guided candidate generation, template-driven feature matching, semantic verification filtering, and spatiotemporal trajectory correction, this paper develops an efficient, low-sample, and highly robust traffic sign detection method, particularly suitable for detecting rare category signs and traffic signs in complex road environments.

[0044] By combining an open vocabulary model, a single-sample recognition algorithm, and Kalman filtering technology, the present invention can provide high-precision, low-error detection solutions in a variety of traffic scenarios, and excels in the detection of rare category signs.

[0045] like Figure 1 As shown, this embodiment provides a traffic sign detection method based on single-sample learning, including the following steps:

[0046] The training, validation, and test sets are constructed based on a traffic sign benchmark dataset. This process includes: removing sparse categories from the dataset to construct a target category set; dividing the target category set into basic categories and novel categories to construct a category distribution system for small-sample learning scenarios; and constructing the training, validation, and test sets based on the category distribution system for small-sample learning scenarios. The specific implementation process includes steps 1 to 3.

[0047] Step 1: Dataset Construction: We selected the publicly available TT100K traffic sign benchmark dataset as the base data source and screened the target categories within the dataset. To ensure the stability and reliability of the evaluation system, we eliminated sparse categories with fewer than 100 samples and constructed a statistically significant set of target categories.

[0048] Step 2: Based on the categories selected in step 1, 10 novel categories under different groups are selected, such as Figure 1 As shown in the figure, the remaining categories are used as basic categories to construct a category distribution system for small sample learning scenarios.

[0049] Step 3: The filtered TT100K original training set is divided into basic categories and novel categories to construct the training set and validation set, respectively. At the same time, the original TT100K test set is re-partitioned according to the filtered categories to form the final test set.

[0050] A pre-trained open vocabulary object detection model is used to generate traffic sign candidate regions in the image for the training set and the validation set, and an intersection-and-union (IoU) is performed on the traffic sign candidate regions to obtain task-related traffic sign candidate regions and category labels.

[0051] Furthermore, the process of obtaining task-related traffic sign candidate regions and category labels includes: using a pre-trained open vocabulary object detection model and coarse-grained semantic cues on the training set and the validation set to obtain traffic sign candidate regions in the image; and obtaining task-related traffic sign candidate regions and category labels based on the intersection-over-union ratio between the traffic sign candidate regions and the annotated bounding boxes. The specific implementation process includes step 4.

[0052] In step 4, a pre-trained open-vocabulary object detector is used to generate potential traffic sign candidate regions in the image by inputting coarse-grained semantic cues. Subsequently, a threshold is set based on the intersection-over-union (IOU) calculation between the candidate regions and the annotated bounding boxes to identify task-relevant traffic signs. Task-relevant traffic signs that meet the IOU threshold are cropped and stored in corresponding subdirectories according to their corresponding category labels. Task-irrelevant traffic signs that do not meet the IOU threshold are uniformly classified into a separate negative sample directory.

[0053] A single-sample recognition algorithm based on metric learning is used to train the candidate traffic sign regions and category labels related to the task to obtain a trained traffic sign detection model. The specific implementation process includes step 5.

[0054] In step 5, a traffic sign single-sample recognition algorithm based on metric learning is used to train the model for the task-related traffic signs extracted in step 4. During training, model parameter optimization and hyperparameter tuning are performed based on the training data, and the optimal model weights for the validation set are ultimately saved.

[0055] The test image of the test set is processed based on the trained traffic sign detection model to obtain a preliminary matching result. The specific implementation process includes step 6.

[0056] Step 6: In the inference phase, the pre-trained open vocabulary object detection model is first used to locate the potential traffic sign area in the image. Then, the metric learning model trained in Step 5 is used to calculate the minimum distance between the area to be detected and the traffic sign templates in the sample library in the feature space, and the nearest traffic sign template is selected as the preliminary matching result.

[0057] Based on the preliminary matching results, a visual language model and semantic clue guide words are used to obtain a semantic verification result. The specific implementation process includes step 7.

[0058] Step 7: Build a visual semantic verification module based on the visual language model, and use semantic cue guide words to determine whether the potential traffic sign area corresponds to its nearest traffic sign template, thereby filtering out task-irrelevant traffic sign categories.

[0059] Based on the semantic verification result, a Kalman filter and intersection-over-union matching strategy are used to obtain the final traffic sign detection result. The specific implementation process includes step 8.

[0060] In step 8, a Kalman filter is used to construct a target trajectory model, combined with an intersection-over-union (IoU) matching strategy to associate category IDs with candidate regions across frames. By integrating category observations from historical frames, temporal consistency corrections are applied to false positives in the current frame, enhancing the spatiotemporal robustness of the detection system.

[0061] Furthermore, based on the semantic verification result, a process of obtaining the final traffic sign detection result by adopting a Kalman filter and an intersection-in-union matching strategy includes: constructing a target motion trajectory model through Kalman filtering, processing the preliminary matching result based on the target motion trajectory model, and obtaining a target motion trajectory prediction result; based on the target motion trajectory prediction result, adopting an intersection-in-union matching strategy to obtain a category ID association result of a cross-frame candidate area; based on the category ID association result of the cross-frame candidate area, adopting a majority voting mechanism to obtain a false detection correction result of the current frame; based on the false detection correction result of the current frame, adopting integration processing to obtain the final traffic sign detection result.

[0062] The structure of the single-sample traffic sign recognition network framework in step 5 is as follows:

[0063] The framework consists of three parts: (1) Transferable Feature Encoder: This module generates stable feature representations through transfer learning based on the DenseNet121 model pre-trained on ImageNet. (2) Latent Space Encoder: This module maps the features extracted from the transferable feature encoder to an efficient latent space, making the features more discriminative. (3) Minimum Distance Classifier: This module calculates the distance between the query image and the support images through metric learning methods to achieve classification.

[0064] As an optional implementation scheme of this embodiment, the traffic sign samples and their standardized templates in the present invention are selected from the TT100K dataset. The TT100K dataset contains a variety of traffic sign types and is a standard dataset in the field of traffic sign detection, widely used for training and evaluating various traffic sign detection models.

[0065] In the implementation of steps 1 to 3, the TT100K dataset is first screened to ensure that each category contains at least 100 instances. In addition, considering that traffic signs in the "io", "po" and "wo" categories lack obvious image content features, these categories are eliminated from the dataset. Through the above screening process, 42 categories in the TT100K dataset are finally selected for the experiment. Figure 1 The grouping strategy shown selects 10 novel categories under different groupings from these 42 categories and defines the remaining 32 categories as basic categories, so as to ensure that the sample data volume of the selected categories is sufficient, providing the necessary conditions for rigorous performance evaluation and experimental verification.

[0066] The method of the present invention mainly includes four key parts: semantic-guided candidate generation, template-driven feature matching, semantic verification filtering and spatiotemporal trajectory correction. In this method, the core structure of the network module consists of the first three parts mentioned above, such as Figure 2 shown.

[0067] In our method, the semantically guided candidate generation step uses a pre-trained open-vocabulary object detector, GroundingDino. This model uses coarse-grained semantic cues, such as "traffic sign" and "circle sign," as input to generate potential traffic sign candidate regions in an image.

[0068] Template-driven feature matching: This is a traffic sign recognition algorithm SENet based on metric learning. The algorithm mainly consists of three parts. The framework is composed of three parts: (1) Transferable feature encoder: This module is based on the DenseNet121 model pre-trained on ImageNet and generates stable feature representations through transfer learning. (2) Latent space encoder: This module maps the features extracted from the transferable feature encoder to an efficient latent space, making the features more discriminative. (3) Minimum distance classifier: Through metric learning methods, the distance between the query image and the support image is calculated to achieve classification.

[0069] Based on the output of the minimum distance classifier, the probability of each category is calculated. The specific formula is as follows:

[0070]

[0071] Where x is the feature vector of the input image, u k represents the prototype of the kth type of traffic sign, and n represents the total number of prototypes of traffic signs. Finally, the category with the highest probability is determined as the predicted category of the query image, u j represents the prototype of the j-th type of traffic sign, and P(y=k|x) is the predicted probability that the image belongs to the k-th type of traffic sign.

[0072] The mean square error loss function is used to optimize the model. The formula is as follows:

[0073]

[0074] where y i is the one-hot encoding of the true label, is the class probability predicted by the model, and N is the number of samples in the training batch.

[0075] During the training process, the Adam optimizer is used to update the model parameters. The initial learning rate is 1×10^ -4 , the value of epsilon is 1×10^ -8 , ensuring gradient stability and robustness. Training was performed for a total of 300 epochs, with the optimal model selected every 50 epochs. The optimal model was then loaded and the learning rate was halved until training was complete.

[0076] Based on the above steps, combined with the pre-trained open vocabulary object detection model and the metric learning single-sample traffic sign recognition method, three sets of results are obtained as shown in Table 1:

[0077] Table 1

[0078] Grouping Average precision across all categories Average precision of the base class Average precision on novel categories Group 1 33.66％ 35.88％ 26.55％ Group 2 35.78％ 37.53％ 30.16％ Group 3 35.14％ 36.39％ 31.13％

[0079] Table 2 compares this model with other models under the first set of settings. The results are shown in Table 2.

[0080] Table 2

[0081] Model Average precision across all categories Average precision of the base class Average precision on novel categories MetaFasterRcnn 19.87％ 25.49％ 1.90％ AirDet 21.19％ 26.47％ 4.28％ ours 33.66％ 35.88％ 26.55％

[0082] Then the semantic verification filtering module is carried out. After completing the template-driven feature matching, the potential traffic sign area and its corresponding nearest traffic sign template are obtained. In order to further verify whether the area matches the traffic sign template, a visual language model (such as GPT4-O) is used to guide the model judgment. In this process, the guiding words of the model are: "Please analyze the following two pictures: one is the template image of the traffic sign, and the other is the real scene image that the sign may correspond to. It is necessary to strictly compare and judge whether the real image matches the template image from the following three aspects. If any one does not match, the overall conclusion is False: Color attributes: The main color / color scheme (such as red background and white text, blue background and white pattern) must be fully matched; Shape features: Contour geometry (such as circle, triangle, rectangle), border style (solid line / dashed line / with or without border), and the overall proportion of the pattern (aspect ratio, relative position of each element) must be completely consistent with the template; Core pattern content: Specific symbols in the sign Key information, including numbers, text, and graphic meanings, must be completely consistent (for example, speed limit signs must have identical numerical values, prohibition signs must have identical symbols, and text content and orientation must match exactly). Please output your conclusion strictly in the following JSON format. The "Analysis" field must describe the comparison results for each of the three aspects (including the specific details of the match / mismatch), and the "Conclusion" field must be a Boolean value (outputting True only if all three aspects match exactly, otherwise False). This JSON output facilitates scripting to obtain binary classification results. A False output will filter out the potential traffic sign area.

[0083] The spatiotemporal trajectory correction module is primarily used in video sequence detection scenarios. It constructs a target motion trajectory model using Kalman filtering and predicts the target's position in each frame. Based on an intersection-over-union (IoU) matching strategy, it calculates the spatial overlap between candidate regions in the current frame and historical trajectories. A threshold is set to filter out areas of high overlap, enabling cross-frame category ID association. A sliding window of five frames is used to store historical category predictions for the same ID. A majority vote mechanism selects the most frequently occurring category and corrects isolated false positives in the current frame, effectively enhancing the spatiotemporal consistency and robustness of detection results in dynamic environments. Figure 3 A flowchart of the spatiotemporal trajectory correction module is shown.

[0084] The single-sample traffic sign detection algorithm proposed in this paper constructs a closed-loop detection process of "semantically guided candidate generation - template-driven feature matching - semantic verification filtering - spatiotemporal trajectory correction". It achieves accurate positioning and recognition of targets in images through a standardized traffic sign template library: first, relying on a pre-trained open vocabulary model, candidate regions containing semantic priors are generated. Then, through metric learning, the initial distance between the region to be detected and the template library is calculated in the feature space to achieve fast matching. The visual language model is then used to verify the semantic alignment between the candidate region and the target template to filter out interference. Finally, a Kalman filter combined with an intersection-over-union matching strategy is used for video sequences to achieve cross-frame ID association and false detection correction. Table 2 verifies that the algorithm significantly outperforms other single-sample traffic sign methods in category detection, indicating its strong application potential in the field of traffic sign detection.

[0085] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A traffic sign detection method based on single-sample learning, characterized in that: The following steps are involved: Construct training, validation, and test sets based on the traffic sign benchmark dataset; Generate traffic sign candidate regions in the image using a pre-trained open vocabulary object detection model for the training set and the validation set, and perform intersection-over-union (IoU) screening on the traffic sign candidate regions to obtain task-related traffic sign candidate regions and category labels; A single-sample recognition algorithm based on metric learning is used to train the candidate traffic sign regions and category labels related to the task to obtain a trained traffic sign detection model; Processing the test image of the test set based on the trained traffic sign detection model to obtain a preliminary matching result; Based on the preliminary matching results, the visual language model and semantic clue guide words are used to obtain the semantic verification results: Based on the semantic verification result, a Kalman filter and intersection-over-union matching strategy is adopted to obtain the final traffic sign detection result.

2. The method according to claim 1, wherein: The process of constructing the training set, validation set, and test set based on the traffic sign benchmark dataset includes: Eliminating sparse categories in the traffic sign benchmark dataset to construct a target category set; Dividing the target class set into basic categories and novel categories, and constructing a category distribution system for small sample learning scenarios; The training set, validation set, and test set are constructed based on the category distribution system in the small sample learning scenario.

3. The method according to claim 1, wherein: The process of obtaining task-related traffic sign candidate regions and category labels includes: For the training set and the validation set, using a pre-trained open vocabulary object detection model and coarse-grained semantic cue words to obtain candidate traffic sign regions in the image; Based on the intersection-over-union ratio between the traffic sign candidate region and the labeled box, the task-related traffic sign candidate region and category label are obtained.

4. The method according to claim 3, wherein: The trained traffic sign detection model is a single-sample traffic sign recognition network; The framework of the single-sample traffic sign recognition network includes: a transferable feature encoder, a latent space encoder and a minimum distance classifier.

5. The method according to claim 4, characterized in that: The minimum distance classifier calculates the distance between the query image and the support image through a metric learning method to achieve classification and calculate the probability of each category; Among them, the expression for calculating the probability of each category is: Among them, x is the feature vector of the input image, u k represents the prototype of the kth type of traffic sign, n represents the total number of prototypes of traffic signs, u j represents the prototype of the j-th type of traffic sign, and P(y=k|x) is the predicted probability that the image belongs to the k-th type of traffic sign.

6. The method according to claim 1, wherein: The process of obtaining a preliminary matching result based on the trained traffic sign detection model includes: Based on the trained traffic sign detection model, the minimum distance between the area to be detected and the traffic sign template is obtained; Based on the minimum distance, a preliminary matching result is obtained.

7. The method according to claim 1, wherein: Based on the semantic verification results, the process of obtaining the final traffic sign detection results by using a Kalman filter and intersection-over-union matching strategy includes: constructing a target motion trajectory model through Kalman filtering, processing the preliminary matching result based on the target motion trajectory model, and obtaining a target motion trajectory prediction result; Based on the target motion trajectory prediction result, an intersection-over-union matching strategy is adopted to obtain the category ID association result of the cross-frame candidate region; The category ID association results based on the candidate regions across frames are processed using a majority voting mechanism to obtain the false detection correction result of the current frame; The false detection correction results based on the current frame are integrated to obtain the final traffic sign detection result.

Images