Small sample target detection system and method adaptive to airport complex scene

Abstract

The invention relates to a small sample target detection system and method adaptive to an airport complex scene. The system comprises the steps that an image trunk feature extraction module extracts a multi-scale semantic feature map of an input image; the text embedding and coding module is used for coding and extracting an input text representing a category to obtain a semantic embedding vector; a category perception convolution kernel construction module extracts local visual features from the mesoscale feature map, and weights the local visual features to generate a category perception convolution kernel; a sliding convolution region matching module calculates the response intensity of each position and category perception convolution kernel in the multi-scale semantic feature map, and determines a center point based on the response intensity; the precise positioning module extracts a plurality of low-confidence threshold candidate frames to construct a candidate frame set, and selects the candidate frame closest to the center point from the candidate frame set as a target detection result; small target detection can be carried out by considering the detection speed, the positioning precision and the semantic generalization ability under the scene that samples are scarce and category features are easy to confuse.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and target detection technology, and in particular to an adaptive small-sample target detection system and method for complex airport scenes. Background Technology

[0002] In practical applications such as airport security, baggage monitoring, and aircraft scheduling, there is often a need to rapidly introduce the ability to detect new types of targets, such as new equipment, specific vehicles, or prohibited items. However, such new targets often lack a large number of labeled samples, and the actual environment is complex with high similarity between targets. Figure 1 The diagram illustrates various types of items with high semantic similarity, posing a significant challenge to existing detection models. Currently widely used detection models, such as the YOLO series, Faster R-CNN, and Transformer-based DETR, typically rely on predefined category sets and large amounts of labeled data, making it difficult to directly adapt to the detection needs of new categories.

[0003] To improve the generalization ability of models, various technical approaches have been proposed. One type is based on transfer learning or meta-learning, which fine-tunes a pre-trained model on a small number of samples of a new category to adapt it to the new task. However, these methods are often computationally expensive, the fine-tuning process is complex, and they struggle to handle situations with high semantic similarity between categories. Another type is Open Vocabulary Detection (OVD), which uses natural language descriptions to define new categories and leverages visual-language models to achieve cross-category generalization. Representative works include OWL-ViT, GLIP, and YOLO-World. Although these methods have the potential for zero-shot detection, they still have significant limitations: First, they are highly dependent on the accuracy of language descriptions. When the visual features of targets are similar and difficult to distinguish precisely using text, semantic confusion and mismatches are prone to occur. Second, in terms of spatial localization accuracy, especially for small or edge targets, existing methods are insufficient in regressing bounding boxes. In addition, some advanced models have complex structures and high computational costs, making them difficult to deploy in scenarios with high real-time requirements, such as airports. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an adaptive small-sample target detection system and method for complex airport scenarios. It provides a small-sample target detection scheme that can balance detection speed, positioning accuracy, and semantic generalization ability in airport scenarios where samples are scarce and category features are easily confused, so as to achieve rapid, accurate, and stable identification of new targets.

[0005] According to a first aspect of the present invention, an adaptive small-sample target detection system for complex airport scenes is provided, comprising: an image backbone feature extraction module, a text embedding encoding module, a category-aware convolutional kernel construction module, a sliding convolutional region matching module, and a precise localization module; The image backbone feature extraction module extracts a multi-scale semantic feature map of the input image; The text embedding encoding module encodes and extracts semantic embedding vectors from the input text representing the category; The category-aware convolutional kernel construction module extracts local visual features from the mid-scale feature map in the multi-scale semantic feature map, calculates the semantic response intensity of each channel of the local visual features based on the semantic embedding vector to obtain the channel attention coefficient, and generates a category-aware convolutional kernel by weighting the local visual features based on the channel attention coefficient. The sliding convolution region matching module calculates the response intensity of each position in the multi-scale semantic feature map with the category-aware convolution kernel, and determines the center point based on the magnitude of the response intensity; The precise positioning module extracts multiple low-confidence threshold candidate boxes to construct a candidate box set, and selects the candidate box closest to the center point from the candidate box set as the target detection result.

[0006] Based on the above technical solution, the present invention can also be improved as follows.

[0007] Optionally, the image backbone feature extraction module is the backbone network in the YOLO-World network; the formula for the multi-scale semantic feature map output by the image backbone feature extraction module is: ; in, Let i be the semantic feature map at the i-th scale. For channel dimension, , The first The height and width of the layer feature map.

[0008] Optionally, the text embedding encoding module further includes: mapping the semantic embedding vector to a space with consistent channel dimensions through a linear projection operation; The semantic embedding vector after mapping is obtained. ; in, and These are the semantic embedding vectors before and after the mapping. , Let C be the new category to be detected, TextEncoder be the encoder extraction operation, and C be the number of categories to be detected in the current task. Indicates the text embedding dimension; , These are learnable parameters.

[0009] Optionally, the process by which the category-aware convolutional kernel construction module extracts local visual features from the mesoscale feature map includes: From the labeled new category sample images, based on the ground truth bounding box coordinates, in the mesoscale feature map Local visual features were extracted using RoIAlign. ; in, , To fix the template size, Indicate category Local visual features in image space.

[0010] Optionally, the formula for calculating the channel attention coefficients obtained by the category-aware convolutional kernel construction module is as follows: ; in, The channel attention coefficient represents the strength of a channel's response to semantics. This indicates global average pooling of the channel. It is the Sigmoid activation function. Indicate category Local visual features in image space.

[0011] Optionally, the formula for generating the category-aware convolutional kernel by the category-aware convolutional kernel construction module is as follows: .

[0012] Optionally, the formula for determining the center point by the sliding convolution region matching module is: ; ; in, This indicates the coordinates of the center point to be determined; Indicates position The characteristics of the receptive field For category-aware convolution kernels.

[0013] Optionally, the process by which the precise positioning module constructs the candidate box set includes: Lower the detector confidence threshold and extract multiple candidate boxes from the original detector head of the YOLO-World backbone network to obtain the candidate box set.

[0014] Optionally, the formula used by the precise positioning module to calculate the distance between each candidate box in the candidate box set and the center point is: ; in, This represents the distance between the j-th candidate box and the center point. The center point Mapping coordinates from the feature map scale back to the original image. This represents the center point of the j-th candidate box.

[0015] According to a second aspect of the present invention, an adaptive method for small-sample target detection in complex airport scenarios is provided, comprising: Step 1: Extract the multi-scale semantic feature map of the input image; Step 2: Extract the semantic embedding vector from the input text representing the category; Step 3: Extract local visual features from the mesoscale feature map in the multi-scale semantic feature map; calculate the semantic response intensity of each channel of the local visual features based on the semantic embedding vector to obtain the channel attention coefficient; and generate a category-aware convolution kernel by weighting the local visual features based on the channel attention coefficient. Step 4: Calculate the response intensity of each position in the multi-scale semantic feature map with the category-aware convolutional kernel, and determine the center point based on the magnitude of the response intensity; Step 5: Extract multiple low-confidence threshold candidate boxes to construct a candidate box set, and select the candidate box closest to the center point from the candidate box set as the target detection result.

[0016] This invention provides an adaptive few-sample target detection system and method for complex airport scenarios. First, a category-aware convolutional kernel construction module is proposed. This module generates semantic templates for new categories through image-text fusion, guiding the detection branch to generate semantically consistent convolutional kernels, thus enhancing the model's ability to discriminate new categories and its adaptability to few samples. Second, to address the problem of inaccurate target localization for new categories, this invention further introduces a precise localization module. This module performs spatial geometric matching between the response region obtained from sliding convolution and a large number of low-confidence candidate boxes, achieving precise localization of new category targets. Finally, to improve detection speed, this invention designs an efficient region matching module based on sliding convolution. This module uses image feature semantic templates as convolutional kernels and slides across the image feature map to calculate response values, thereby quickly completing candidate region selection without requiring region-by-region comparison, greatly improving detection efficiency. The three-stage detection method proposed in this invention systematically solves the core problems of YOLO-World in few-sample new category detection scenarios, possessing good semantic generalization ability, real-time detection performance, and target localization accuracy. Attached Figure Description

[0017] Figure 1This is a diagram illustrating various types of items that exhibit high semantic similarity. Figure 2 A flowchart illustrating an embodiment of an adaptive small-sample target detection system for complex airport scenarios provided by the present invention; Figure 3 This is a schematic diagram illustrating the actual deployment effect of an embodiment of the adaptive small-sample target detection system for complex airport scenarios provided by the present invention. Detailed Implementation

[0018] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0019] Figure 2 A flowchart illustrating an embodiment of the adaptive small-sample target detection system for complex airport scenarios provided by the present invention is shown below. Figure 2 As shown, the small sample target detection system includes: an image backbone feature extraction module, a text embedding encoding module (CLIP Text Encoder), a category-aware convolutional kernel construction module (CCKCM), a sliding convolutional region matching module (ERMM, Efficient Region Matching Module), and a precise localization module (PLM, PreciseLocalization Module).

[0020] The image backbone feature extraction module extracts multi-scale semantic feature maps of the input image.

[0021] The text embedding encoding module encodes and extracts semantic embedding vectors from the input text representing the category.

[0022] The category-aware convolutional kernel construction module extracts local visual features from the mid-scale feature map in the multi-scale semantic feature map. Based on the semantic embedding vector, it calculates the semantic response intensity of each channel of the local visual features to obtain the channel attention coefficient. Based on the channel attention coefficient, the local visual features are weighted to generate the category-aware convolutional kernel.

[0023] The sliding convolutional region matching module calculates the response intensity of each location in the multi-scale semantic feature map with the category-aware convolutional kernel, and determines the center point based on the magnitude of the response intensity.

[0024] The precise localization module extracts multiple low-confidence threshold candidate boxes to construct a candidate box set, and selects the candidate box closest to the center point from the candidate box set as the target detection result.

[0025] YOLO-World, a recently proposed efficient open-vocabulary detection model, combines visual language pre-training techniques with the real-time advantages of the YOLO architecture, enabling it to detect new categories specified in text without fine-tuning. However, it still suffers from detection bias and accuracy degradation when handling challenges such as high-similarity categories, small object localization, and rapid adaptation with few samples.

[0026] To address the shortcomings of the YOLO-World model in few-shot object detection tasks for new categories, such as insufficient semantic generalization and low detection accuracy, this invention proposes a few-shot object detection method based on category-aware convolutional kernels. The overall method comprises three key stages, corresponding to the three detection stages of "feature modeling—region retrieval—bounding box regression," forming an end-to-end object detection workflow for new categories. First, a category-aware convolutional kernel construction module is proposed. This module generates semantic templates for new categories through image-text fusion, guiding the detection branch to generate semantically consistent convolutional kernels, thus enhancing the model's discriminative ability for new categories and its adaptability to few-shot detection. Second, to solve the problem of inaccurate localization of new category objects, this invention further introduces a precise localization module. This module performs spatial geometric matching between the response region obtained from sliding convolution and a large number of low-confidence candidate boxes, achieving precise localization of new category objects. Finally, to improve detection speed, this invention designs an efficient region matching module based on sliding convolution. This module uses the image feature semantic template as the convolutional kernel, sliding across the image feature map to calculate the response value, thereby quickly completing candidate region selection without requiring region-by-region comparison, greatly improving detection efficiency. The three-stage detection method proposed in this invention systematically solves the core problem of YOLO-World in the scenario of detecting new categories with few samples, and has good semantic generalization ability, real-time detection performance and target localization accuracy.

[0027] Example 1 Embodiment 1 provided by this invention is an embodiment of an adaptive small-sample target detection system for complex airport scenarios provided by this invention, combined with... Figure 2 It can be seen that the embodiments of this small sample target detection system include: The system includes an image backbone feature extraction module, a text embedding encoding module, a category-aware convolution kernel construction module, a sliding convolution region matching module, and a precise localization module.

[0028] The image backbone feature extraction module extracts multi-scale semantic feature maps of the input image.

[0029] In one possible embodiment, the image backbone feature extraction module is the backbone network of the YOLO-World network; the formula for the multi-scale semantic feature map output by the image backbone feature extraction module is: .

[0030] in, Let i be the semantic feature map at the i-th scale. For channel dimension, , The first The height and width of the layer feature map.

[0031] In the embodiments provided by this invention, after the input image I is processed by the YOLO-World backbone network, three sets of semantic feature maps with different resolutions are output. These are used to cover receptive fields of targets at different scales. In this embodiment of the invention, a mesoscale feature map is used. As the foundational visual representation constructed from category-aware convolutional kernels, it aims to achieve a balance between computational cost and semantic information.

[0032] In its implementation, the YOLO-World network can be based on YOLOv8 and uses Darknet (the open-source neural network framework Darknet-53) as its image encoder. Darknet is a deep convolutional neural network capable of extracting multi-scale features from images. YOLO-World constructs a Feature Pyramid (FPN) to fuse feature maps of different scales, enhancing the model's ability to detect targets of different sizes.

[0033] The text embedding encoding module encodes and extracts semantic embedding vectors from the input text representing the category.

[0034] In one possible embodiment, the text embedding encoding module further includes mapping the semantic embedding vector to a channel-dimension consistent space through a linear projection operation.

[0035] Obtain the mapped semantic embedding vector .

[0036] in, and These are the semantic embedding vectors before and after the mapping. , Let C be the new category to be detected, TextEncoder be the encoder extraction operation, and C be the number of categories to be detected in the current task. Indicates the text embedding dimension; , These are learnable parameters.

[0037] In practice, the text embedding encoding module uses CLIP's TextEncoder to convert each target category (or natural language description) into a fixed-length semantic embedding: .

[0038] Specifically, the pre-trained CLIP text encoder encodes the input text (such as category names, noun phrases, or object descriptions) into text embeddings. CLIP is a vision-language pre-trained model that maps text and image features to the same semantic space.

[0039] To achieve consistent alignment with image features, this invention introduces a linear projection operation to map the text vector to a space with consistent channel dimensions, making the output... Consistent with the dimension of image feature channels.

[0040] The category-aware convolutional kernel construction module extracts local visual features from the mid-scale feature map in the multi-scale semantic feature map. Based on the semantic embedding vector, it calculates the semantic response intensity of each channel of the local visual features to obtain the channel attention coefficient. Based on the channel attention coefficient, the local visual features are weighted to generate the category-aware convolutional kernel.

[0041] To address the issue of the YOLO-World model lacking explicit semantic guidance for new categories in few-shot object detection scenarios, this invention provides a category-aware convolutional kernel construction module. This module fuses the textual semantic embedding of the new category with the corresponding local visual features of the image to construct a convolutional kernel with category discrimination capabilities. This kernel is then used for subsequent sliding matching operations, thereby enhancing the model's image-text consistency and category awareness, and improving the retrieval accuracy for new category target regions.

[0042] In one possible implementation, the process by which the category-aware convolutional kernel building module extracts local visual features from the mesoscale feature map includes: From the labeled new category sample images, based on the ground truth bounding box coordinates, in the mesoscale feature map Local visual features were extracted using RoIAlign. .

[0043] in, , To fix the template size, Indicate category Local visual features in image space.

[0044] In one possible implementation, the formula for calculating the channel attention coefficients obtained by the category-aware convolutional kernel building module is as follows: .

[0045] in, The channel attention coefficient represents the strength of a channel's response to semantics. This indicates global average pooling of the channel. For the Sigmoid activation function, Indicate category Local visual features in image space.

[0046] In specific implementation, in order to integrate visual features and text semantics, this embodiment of the invention introduces a channel-level attention mechanism to calculate the semantic response intensity of each channel as the channel attention coefficient.

[0047] The formula for generating category-aware convolutional kernels using the category-aware convolutional kernel building module is as follows: .

[0048] In practice, this template will be used as input to the subsequent sliding convolution region matching module with a fixed convolution kernel to locate the response region in the image that is most consistent with the semantics of the new category.

[0049] The CCKCM module introduces a structured image-text fusion mechanism, establishing an explicit semantic guidance mechanism at the convolutional level. Compared to YOLO-World's original method of embedding spatial dot products, it has stronger category awareness and spatial discrimination capabilities. This module has good pluggability and versatility, and can adapt to scenarios where new categories are added incrementally.

[0050] The sliding convolutional region matching module calculates the response intensity of each location in the multi-scale semantic feature map with the category-aware convolutional kernel, and determines the center point based on the magnitude of the response intensity.

[0051] In one possible embodiment, the formula for the sliding convolution region matching module to determine the center point is: ; .

[0052] in, This indicates the coordinates of the center point to be determined; Indicates position The characteristics of the receptive field For category-aware convolution kernels.

[0053] In specific implementation, to quickly locate the region most similar to the new category semantics in the entire image, this embodiment of the invention is based on a category-aware convolutional kernel. Design a sliding convolution region matching module.

[0054] This sliding convolutional region matching module uses category-aware convolutional kernels. Using the target category template as a reference, the response intensity of each position in the image feature map to the target category perception template is calculated through sliding matching, thereby completing the coarse localization of the candidate region.

[0055] In practice, the response map can be obtained using a standard two-dimensional convolution method: .

[0056] in, This is a mid-scale feature map of the YOLO-World backbone network. For category-aware templates, For response diagrams.

[0057] Each location The response value reflects the degree of matching between the region and the target category perception template. The response value can be interpreted as the cosine similarity between the local region and the target category perception template.

[0058] To improve efficiency, embodiments of the present invention use target category-aware templates. Perform pre-normalization and assume The norm does not change significantly across different locations, thus simplifying the normalization calculation. Finally, the location corresponding to the maximum response value in the response map RRR is selected as the initial matching center for the current category; this location is the match center with the category. semantic templates The center point of the best-matching image region.

[0059] The center point It will be passed to the precise localization module for further filtering of candidate detection boxes that are consistent with the location space, thereby realizing the transition from semantic matching to bounding box regression.

[0060] Although the sliding convolutional region matching module (ERMM) can locate the center point of the region in the image that best matches the target class using the response map. However, this location only provides a coarse-grained response area and does not contain explicit bounding box information. To achieve accurate prediction of the target location and size, this embodiment of the invention provides a precise positioning module that transitions from coarse-matched response to final bounding box regression by judging the geometric spatial consistency between the candidate box set and the response location.

[0061] The precise localization module extracts multiple low-confidence threshold candidate boxes to construct a candidate box set, and selects the candidate box closest to the center point from the candidate box set as the target detection result.

[0062] In one possible embodiment, the process of the precise positioning module constructing the candidate box set includes: Lower the detector confidence threshold and extract multiple candidate boxes from the original detector head of the YOLO-World backbone network to obtain a candidate box set.

[0063] In practice, to ensure that no potential targets are missed, the detector confidence threshold can be lowered from the default 0.25 to 0.001 to obtain a preliminary candidate box set. .

[0064] In one possible embodiment, the formula used by the precise positioning module to calculate the distance between each candidate box in the candidate box set and the center point is: .

[0065] in, This represents the distance between the j-th candidate box and the center point. Center point Mapping coordinates from the feature map scale back to the original image. This represents the center point of the j-th candidate box.

[0066] Match the category peak points obtained from the response graph Mapping back to the original image coordinate system from the feature map scale, denoted as... The system calculates the Euclidean distance between the candidate box and the center point of each candidate box, and finally selects the candidate box closest to the center point of the response as the final detection result.

[0067] The localization strategy of this precise localization module is based on the assumption of spatial location consistency: the model has semantically determined the target region through sliding matching, while candidate boxes provide scale and shape information. The combination of the two can achieve fusion optimization of localization and regression. Since the candidate boxes come from the backbone network and have rich prior boundary information, and the response centers come from semantic matching and have strong semantic relevance, the fusion of the two in spatial geometry can significantly improve the accuracy of the final detection results.

[0068] This precise localization module achieves spatial alignment between the sliding matching result and the output of the detection head through the candidate box set, thereby significantly improving the localization accuracy of the model in scenarios with few samples and new categories.

[0069] This invention provides an adaptive few-sample target detection system for complex airport scenarios. For new categories, this invention introduces a category-aware convolutional kernel construction module, fusing local image region features corresponding to a small number of labeled samples with their textual semantic embeddings to construct a semantically guided convolutional kernel. This template retains the structural information of visual features while incorporating category semantics, which helps improve the model's ability to distinguish new classes under conditions of few samples.

[0070] Next, we move on to the sliding convolution region matching stage. The constructed convolution kernels... As a fixed template, in the image feature map Perform a two-dimensional sliding convolution operation on the top to obtain the response. The response value at each location represents the similarity between that region and the category template. To avoid high computational overhead, this invention employs a normalization approximation strategy, simplifying the convolution calculation to a standard forward propagation process. Finally, the point with the maximum response is selected from the response map. This serves as the initial matching location of the new category target in the image.

[0071] Finally, to transform semantic response locations into precise bounding boxes, this invention introduces a Precise Localization (PLM) module. This module extracts a large number of low-confidence candidate boxes from the backbone detector and transforms the response locations... Mapping to the image coordinate system, the Euclidean distance between the detected bounding box and the center point of all candidate boxes is calculated, and the closest one is selected as the final detection result. This strategy achieves geometric alignment between semantic response and detection structure, balancing expressive power and localization accuracy.

[0072] The overall process is based on template construction, uses convolutional matching as a bridge, and geometric alignment as the final step, realizing a complete closed loop from perceptual modeling to accurate detection of new category targets, and has good semantic generalization, matching efficiency and localization robustness.

[0073] Example 2 Embodiment 2 provided by this invention is an embodiment of an adaptive small-sample target detection method for complex airport scenarios provided by this invention, combined with... Figure 2 It can be seen that the embodiments of this small sample target detection method include: Step 1: Extract the multi-scale semantic feature map of the input image.

[0074] Step 2: Extract the semantic embedding vector from the input text representing the category.

[0075] Step 3: Extract local visual features from the mesoscale feature map in the multi-scale semantic feature map, calculate the semantic response intensity of each channel of the local visual features based on the semantic embedding vector to obtain the channel attention coefficient, and generate a category-aware convolution kernel by weighting the local visual features based on the channel attention coefficient.

[0076] Step 4: Calculate the response intensity of each location in the multi-scale semantic feature map with the category-aware convolutional kernel, and determine the center point based on the magnitude of the response intensity.

[0077] Step 5: Extract multiple low-confidence threshold candidate boxes to construct a candidate box set, and select the candidate box closest to the center point from the candidate box set as the target detection result.

[0078] It is understood that the adaptive small sample target detection method for complex airport scenarios provided by the present invention corresponds to the adaptive small sample target detection system for complex airport scenarios provided in the foregoing embodiments. The relevant technical features of the adaptive small sample target detection method for complex airport scenarios can be referred to the relevant technical features of the adaptive small sample target detection system for complex airport scenarios, and will not be repeated here.

[0079] Example 3 Embodiment 3 provided by the present invention is a specific application embodiment of the adaptive small sample target detection system for complex airport scenarios provided by the present invention, which relates to the fields of computer vision, artificial intelligence, and airport automation, namely, using surveillance cameras and utilizing computer vision technology to identify airport targets.

[0080] To realize the adaptive small-sample target detection system for complex airport scenarios proposed in this invention, the system deployment covers three core parts: cloud training server, edge inference server, and monitoring client, all of which are operated and managed on the computer.

[0081] The cloud-based training server, equipped with high-performance GPUs, receives new category sample images and annotation information uploaded by users via the web interface. On this server, based on a pre-trained YOLO-World model, new category semantic templates are generated and the model is fine-tuned using a category-aware convolutional kernel construction module. Users can log in to the system interface via a browser, submit data, and start the training task. After training is complete, the optimized model parameters and semantic templates are packaged and distributed to the edge inference nodes.

[0082] The edge inference server is deployed at the airport, directly connected to the existing airport monitoring network, and is responsible for real-time processing of video streams. This server integrates three core modules of the detection model (CCKCM, ERMM, and PLM), extracts multi-scale features from the input image, performs sliding convolution matching using semantic templates to quickly locate the target region, and combines geometric constraints to complete accurate bounding box regression. Detection results are pushed to the monitoring terminal in real time and written to a database for log auditing and event backtracking.

[0083] The monitoring client is a browser-accessible web application that provides functions such as user authentication, task management, real-time monitoring, and result visualization. Airport operators can select monitoring points, view real-time detection results, retrieve historical alarm records, and receive abnormal status notifications pushed by the system on their computers. All interactions are completed through a web page, requiring no additional mobile applications or terminal software installation.

[0084] This system supports multi-node, high-concurrency processing and is suitable for large-scale airport monitoring scenarios. It balances real-time detection, accuracy, and system stability. A diagram illustrating the actual deployment effect is shown below. Figure 3 As shown.

[0085] This invention, with its unique technical advantages and design tailored to practical needs, possesses significant practical value and broad industry application potential. Its core algorithm is based on the high-performance YOLO-World structure with targeted improvements. While maintaining high detection speed and ensuring real-time response capabilities, it innovatively incorporates a few-shot learning mechanism, greatly enhancing the ability to quickly identify and accurately locate new types of targets. This characteristic makes it perfectly suited for special scenarios like airports, where target types are extremely diverse and the environment is complex and ever-changing. Whether it's pedestrians, various ground support equipment, or baggage and cargo of different sizes, they can all be efficiently identified and tracked. With the accelerating pace of smart airport construction, the demand for efficient and accurate video intelligent analysis technology is increasing across all aspects of airport operations. This invention precisely meets this market demand and can be widely applied to multiple key business scenarios within airports. In airport security monitoring, it can monitor abnormal personnel or items in real time and provide timely warnings of potential security risks. In ground vehicle dispatching, it can accurately track vehicle locations and operating status, optimizing dispatch efficiency and reducing congestion. In baggage transfer tracking, it can monitor the entire baggage flow path, reducing the risk of loss or misdelivery. In restricted area intrusion detection, it can quickly identify unauthorized personnel or objects entering restricted areas, ensuring the security of core airport areas. Through applications in these scenarios, this invention can effectively improve the overall operational efficiency and security management level of airports, providing strong support for smart airport construction. Furthermore, the system can be fully managed via a computer, without relying on additional mobile devices. This significantly lowers the deployment threshold for airports, simplifies daily usage processes, and facilitates seamless integration with existing airport information systems, reducing the cost and difficulty of system integration. Furthermore, the algorithm module of this invention possesses excellent scalability. Besides stably supporting typical target detection tasks commonly encountered in airports, it can be further integrated with more advanced functions such as behavior recognition and event prediction based on actual needs, continuously expanding its application boundaries in airport scenarios and providing airports with more comprehensive and intelligent solutions. In the long term, this technology can also be easily extended to similar transportation hub scenarios with similar operational characteristics to airports, such as train stations, subway stations, and ports. It has strong cross-scenario replicability and large-scale promotion prospects, and is expected to play an important role in the entire transportation sector, contributing to improving the operational efficiency and safety management level of transportation hubs. During the R&D phase, this invention successfully completed the entire development process of the core algorithm and conducted multiple rounds of rigorous verification tests tailored to the specific characteristics of airport scenarios to ensure stable and reliable technical performance. To comprehensively evaluate the algorithm's performance in real-world scenarios, the R&D team not only constructed a self-built scenario dataset containing various typical airport targets but also collected image data from some real-world airport operations. Based on this, a comprehensive performance evaluation of the system was conducted. Experimental results clearly demonstrate that even under extremely small sample conditions with only 1-5 labeled samples per target type, the algorithm still exhibits excellent detection performance for key airport targets such as aircraft, ground support vehicles, and equipment containers. The mean average accuracy (mAP) remains consistently above 85%, accurately identifying the features and locations of various targets. Simultaneously, the algorithm's detection speed exceeds 25 FPS, fully meeting the data processing speed requirements of airport real-time monitoring and dynamic scheduling scenarios. Its overall performance significantly outperforms current mainstream small-sample detection baseline algorithms in the industry, fully demonstrating the superiority and practicality of this technology under small-sample conditions. Currently, based on mature algorithms and system architecture, this invention has been officially piloted in a ground vehicle monitoring scenario at a major international airport in China, bringing the technology from the laboratory to the real airport operation environment. During the pilot phase, the system seamlessly integrates with the airport's existing monitoring network, enabling real-time dynamic detection and precise positioning of various types of ground support vehicles, such as baggage carts, aircraft guidance vehicles, and refueling trucks, without requiring large-scale modifications to existing hardware. With this system, ground dispatchers can monitor the real-time trajectory, location distribution, and current operating status of various vehicles. This allows for more scientific planning of vehicle routes and optimization of dispatching schemes, effectively avoiding problems such as vehicle congestion and route conflicts. It also enables timely detection and warnings of abnormal vehicle stops and violations, providing strong support for ground operation safety. During the pilot testing period, the system operated stably overall, without major malfunctions or performance fluctuations, and maintained a consistently high detection accuracy rate. Positive feedback from airport staff participating in the pilot further validated the effectiveness, practicality, and compatibility of the technology with the existing operational system in a real airport environment, laying a solid foundation for its subsequent widespread adoption.

[0086] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0087] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An adaptive small-sample target detection system for complex airport scenarios, characterized in that, The small sample target detection system includes: an image backbone feature extraction module, a text embedding encoding module, a category-aware convolutional kernel construction module, a sliding convolutional region matching module, and a precise localization module; The image backbone feature extraction module extracts a multi-scale semantic feature map of the input image; The text embedding encoding module encodes and extracts semantic embedding vectors from the input text representing the category; The category-aware convolutional kernel construction module extracts local visual features from the mid-scale feature map in the multi-scale semantic feature map, calculates the semantic response intensity of each channel of the local visual features based on the semantic embedding vector to obtain the channel attention coefficient, and generates a category-aware convolutional kernel by weighting the local visual features based on the channel attention coefficient. The sliding convolution region matching module calculates the response intensity of each position in the multi-scale semantic feature map with the category-aware convolution kernel, and determines the center point based on the magnitude of the response intensity; The precise positioning module extracts multiple low-confidence threshold candidate boxes to construct a candidate box set, and selects the candidate box closest to the center point from the candidate box set as the target detection result.

2. The small sample target detection system according to claim 1, characterized in that, The image backbone feature extraction module is the backbone network in the YOLO-World network; the formula for the multi-scale semantic feature map output by the image backbone feature extraction module is: ； in, Let i be the semantic feature map at the i-th scale. For channel dimension, , The first The height and width of the layer feature map.

3. The small sample target detection system according to claim 1, characterized in that, The text embedding encoding module further includes: mapping the semantic embedding vector to a space with consistent channel dimensions through a linear projection operation; The semantic embedding vector after mapping is obtained. ; in, and These are the semantic embedding vectors before and after the mapping. , Let C be the new category to be detected, TextEncoder be the encoder extraction operation, and C be the number of categories to be detected in the current task. Indicates the text embedding dimension; , These are learnable parameters.

4. The small sample target detection system according to claim 1, characterized in that, The process by which the category-aware convolutional kernel construction module extracts local visual features from the mesoscale feature map includes: From the labeled new category sample images, based on the ground truth bounding box coordinates, in the mesoscale feature map Local visual features were extracted using RoIAlign. ; in, , To fix the template size, Indicate category Local visual features in image space.

5. The small sample target detection system according to claim 3, characterized in that, The formula for calculating the channel attention coefficients obtained by the category-aware convolutional kernel construction module is as follows: ； in, The channel attention coefficient represents the strength of a channel's response to semantics. This indicates global average pooling of the channel. It is the Sigmoid activation function. Indicate category Local visual features in image space.

6. The small sample target detection system according to claim 5, characterized in that, The formula for generating the category-aware convolutional kernel by the category-aware convolutional kernel construction module is as follows: 。 7. The small sample target detection system according to claim 5, characterized in that, The formula for determining the center point by the sliding convolution region matching module is: ； ； in, This indicates the coordinates of the center point to be determined; Indicates position The characteristics of the receptive field For category-aware convolution kernels.

8. The small sample target detection system according to claim 1, characterized in that, The process by which the precise positioning module constructs the candidate box set includes: Lower the detector confidence threshold and extract multiple candidate boxes from the original detector head of the YOLO-World backbone network to obtain the candidate box set.

9. The small sample target detection system according to claim 1, characterized in that, The formula used by the precise positioning module to calculate the distance between each candidate box in the candidate box set and the center point is: ； in, This represents the distance between the j-th candidate box and the center point. The center point Mapping coordinates from the feature map scale back to the original image. This represents the center point of the j-th candidate box.

10. An adaptive method for small-sample target detection in complex airport scenarios, characterized in that, The small sample target detection method includes: Step 1: Extract the multi-scale semantic feature map of the input image; Step 2: Extract the semantic embedding vector from the input text representing the category; Step 3: Extract local visual features from the mesoscale feature map in the multi-scale semantic feature map; calculate the semantic response intensity of each channel of the local visual features based on the semantic embedding vector to obtain the channel attention coefficient; and generate a category-aware convolution kernel by weighting the local visual features based on the channel attention coefficient. Step 4: Calculate the response intensity of each position in the multi-scale semantic feature map with the category-aware convolutional kernel, and determine the center point based on the magnitude of the response intensity; Step 5: Extract multiple low-confidence threshold candidate boxes to construct a candidate box set, and select the candidate box closest to the center point from the candidate box set as the target detection result.

Images