Large model cooperative anil target identification method and system for camera icing fog

Abstract

The application relates to a large model cooperative ANIL target identification method and system for camera icing and fogging, and belongs to the technical field of computer vision and road safety. The application adopts an end-cloud cooperative architecture, dynamically adjusts a dark channel defogging coefficient and a gamma correction coefficient according to visibility and icing degree on the end side, and performs adaptive pretreatment on an image; real-time target detection is performed by using an ANIL light-weight network, and a confidence threshold is dynamically judged based on the visibility; when the confidence is insufficient, a cloud-side Grounding DINO 2.0 large model is triggered to perform accurate reasoning; meanwhile, the large model output is used as prior knowledge to perform online transfer learning and parameter optimization on the end-side ANIL network, thereby forming a continuous evolution closed loop; and finally, millisecond-level priority early warning is realized for a sudden condition. The application is used for target identification and safety early warning of a camera in an intelligent traffic system under bad weather such as icing and fogging, and can improve the identification precision, real-time performance and environmental adaptability of traffic monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical fields of computer vision and road safety, and relates to a large model collaborative ANIL target recognition method and system for camera icing and fogging. Background Art

[0002] Based on the current information and preliminary analysis, this application relates to the vision detection technology in the intelligent transportation system (ITS), especially a method and system for target recognition that integrates a large model and a lightweight network for cameras in harsh weather conditions such as icing and fogging.

[0003] Currently, most traffic camera target recognition systems use conventional convolutional neural networks (CNNs) or their lightweight variants (such as the ANIL network) as the core detection algorithm. The basic architecture of such systems includes an image acquisition module, a single inference module, a simple data storage module, and an output module. Its working principle is as follows: After the camera captures an image, the edge-side inference module directly performs feature extraction and target detection, and stores or uploads the results to the traffic control center. However, in practical applications, especially in the cold and foggy regions in the north of China, the above technical solutions have the following technical defects: First, it is difficult to balance the detection accuracy and real-time performance in harsh weather. Icing and fogging weather can cause image features to be blurred and the contrast to decrease. The traditional CNN network lacks a targeted feature enhancement design, and although the lightweight ANIL network can achieve edge-side real-time inference, its model capacity is limited, lacking prior knowledge for accurate target positioning, and its generalization ability is weak in complex scenarios, resulting in high false negative and false positive rates.

[0004] Second, the edge-cloud collaboration mechanism has defects, restricting the further improvement of system performance. Existing technologies such as Chinese Patent CN120148276B disclose an edge-cloud collaborative intelligent vehicle driving control method, where the edge box performs target detection and the cloud platform performs regional linkage warning. However, the functions of its edge-side model (such as the traffic flow prediction model) and the cloud-side model (the warning judgment model) are different, and there is no "edge-side preliminary screening - cloud-side refined judgment" collaboration for the same target detection task. At the same time, its edge-cloud interaction logic is simple, lacking an adaptive confidence judgment mechanism, which is prone to ineffective uploads and increases latency. In addition, the solutions represented by Chinese Patent CN120932384A place all perception and recognition tasks on the vehicle side, belonging to a "pure edge-side" architecture and failing to utilize the powerful capabilities of the cloud-side large model.

[0005] Furthermore, existing technologies have failed to effectively leverage the advantages of large models to address the limitations of small models in specific scenarios. Large models with cross-modal semantic understanding and accurate target localization capabilities have demonstrated superior performance in general domains, but they have not yet been applied to target detection in traffic camera scenarios such as icy and foggy conditions to compensate for the insufficient generalization ability of lightweight models.

[0006] In addition, existing systems have significant shortcomings in environmental adaptability and model evolution capabilities.

[0007] (1) The preprocessing parameters for icing and fog images are usually fixed values ​​and cannot be dynamically adjusted according to the severity of the weather, resulting in poor enhancement effect. Although existing technologies have proposed dynamic compensation of meteorological parameters, the compensation target is the macroscopic traffic flow prediction model, rather than the preprocessing enhancement of the camera input image.

[0008] (2) Once the edge model is deployed, its parameters are fixed and cannot be continuously optimized using new data generated during operation. The system has poor long-term environmental adaptability and high maintenance costs.

[0009] (3) A multi-dimensional hierarchical data architecture for assisting detection algorithms has not been constructed. Geographic, environmental, and traffic flow data are stored independently, which cannot provide effective contextual information for detection algorithms to improve robustness.

[0010] (4) The lack of priority warning design for sudden situations such as emergency braking and illegal lane changes makes it difficult to meet the safety requirements of intelligent transportation for millisecond-level response.

[0011] Therefore, how to significantly improve the target recognition accuracy of cameras in adverse weather conditions such as icing and fog while ensuring the real-time performance of edge detection, and how to achieve dynamic adaptation of the system to environmental changes and continuous evolution of the model, are technical problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0012] In view of this, the present invention addresses the above-mentioned technical problems by providing a large-model collaborative ANIL target recognition method and system for camera icing and fogging.

[0013] A large-scale collaborative ANIL target recognition method for camera icing and fogging includes the following steps: S1: Multi-source data acquisition, acquiring traffic images captured by road cameras, visibility and camera icing levels captured by environmental sensors, and vehicle status data captured by vehicle-mounted terminals.

[0014] S2: Based on the visibility and camera icing, dynamically adjust the dark channel prior dehazing coefficient ω and gamma correction coefficient γ to perform adaptive preprocessing on the original image to obtain an enhanced image.

[0015] The dark channel prior defogging coefficient ω is dynamically adjusted according to visibility: when visibility < 200m, ω = 0.95; when 200m ≤ visibility ≤ 500m, ω = 0.85; when visibility > 500m, ω = 0.75; the gamma correction coefficient γ is dynamically adjusted according to the icing degree of the camera: when the icing degree > 80%, γ = 2.0; when 50% ≤ icing degree ≤ 80%, γ = 1.5; when the icing degree < 50%, γ = 1.2.

[0016] The calculation formula for the transmittance t(x) of the dark channel prior defogging is:

[0017] where c is the RGB color channel, Ω(x) is the local window centered on x, is the global atmospheric light value of each channel, represents the pixel gray value of any pixel point y in the local window Ω(x) centered on the pixel point x in the original foggy image under the RGB color channel c, and the value range is usually 0~255 (8-bit image); The clear image is restored by:

[0018] where, is the clear image after defogging, is the lower limit of the transmittance, is the original foggy image, is the transmittance, is the global atmospheric light value.

[0019] S3: Use the on-device ANIL network to perform object detection on the enhanced image, and output the object category, object localization box, and on-device detection confidence .

[0020] S4: Determine the dynamic threshold T according to the visibility. If ≥T, then directly output the on-device detection result to the warning module.

[0021] If <T, then upload the original image, enhanced image, and on-device detection result to the cloud; the dynamic threshold T is determined according to the visibility: when visibility > 500m, T = 0.85; when 200m ≤ visibility ≤ 500m, T = 0.90; when visibility < 200m, T = 0.95.

[0022] S5: Use the cloud-based Grounding DINO 2.0 large model to perform cross-modal semantic understanding and precise object localization on the uploaded data, and output the precise object category, refined localization box, and cloud detection confidence .

[0023] End-side weights are determined based on visibility. and cloud weight And calculate the confidence level of edge-cloud fusion.

[0024] Among them, when visibility is >500m, =0.8, =0.2; when 200m ≤ visibility ≤ 500m, =0.6, =0.4; when visibility <200m, =0.4, =0.6.

[0025] The cross-modal attention interaction of the Grounding DINO 2.0 large model is as follows:

[0026] in, Visual features extracted for the visual Transformer For text query features, Text query features The transpose of the matrix, The ViT visual features are obtained after weight mapping, where W represents the weights and T represents the visual transformer. This is the scaling factor.

[0027] S6: Using the refined localization boxes and semantic labels output by Grounding DINO 2.0 as prior knowledge, perform online transfer learning on the edge ANIL network to optimize the convolutional kernel weights W and bias b of the ANIL network, and then send the optimized parameters down to the edge.

[0028] The optimization objective of the online transfer learning is to minimize the loss function of the ANIL network. Where the weight coefficient λ=0.5, and the classification loss Cross-entropy loss, regression loss To smooth out the L1 loss.

[0029] The edge-side ANIL network employs depthwise separable convolutions to construct a lightweight feature extractor, and its feature extraction is as follows:

[0030] Where X is the preprocessed enhanced image. For depthwise separable convolution, the Swish activation function is defined as Swish(x) = x·σ(βx), with β = 1.

[0031] Preferably, the method of the present invention further includes S7: storing the collected environmental data, traffic detection data, vehicle self-data and edge-cloud detection results into the corresponding databases of the layered data storage module, wherein the layered data storage module includes a geographic information layer database, an environmental information layer database, a traffic information layer database, a vehicle self-detection information layer database, and a management information layer database.

[0032] Preferably, the method of the present invention further includes S8: when the end-cloud detection result determines that the vehicle brakes suddenly or changes lanes illegally, the situation is marked as high priority, and a warning signal is sent to the traffic control terminal, the surrounding vehicle warning terminal and the road information broadcasting terminal simultaneously through the output warning module, with a warning response time ≤200ms.

[0033] Furthermore, this invention also provides a large-model collaborative ANIL target recognition system for cameras experiencing icing and fogging, employing an edge-cloud collaborative architecture, comprising: The data acquisition module includes a road camera acquisition unit, a multi-dimensional environmental sensor, and a vehicle-mounted terminal data receiving unit, which is used to execute S1; The edge computing module, deployed on the roadside, includes a microprocessor, an ANIL network inference unit, local temporary memory, and an image preprocessing unit, and is used to execute S2 to S4. The wireless transmitter / receiver module uses the 5G communication protocol to achieve high-speed data transmission between the device and the cloud. The cloud service module includes the Grounding DINO 2.0 large model inference center, cloud servers, and online transfer learning units for executing S5 to S6; The data tiered storage module, deployed in the cloud, includes a geographic information layer database, an environmental information layer database, a traffic information layer database, a vehicle self-detection information layer database, and a management information layer database. The output warning module includes a traffic control terminal, a surrounding vehicle warning terminal, and a road information broadcasting terminal, which are used to output detection results and warnings of emergencies.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Significantly improved accuracy and real-time performance in severe weather identification. By synergistically fusing the Grounding DINO 2.0 large model with the ANIL lightweight network and combining it with an adaptive image preprocessing algorithm, the detection rate of traffic targets in icy and foggy scenarios was improved, and the false negative and false positive rates were significantly reduced. The edge-side ANIL network meets the 30fps real-time inference requirement. The visibility-based dynamic weight fusion mechanism only triggers cloud inference when the confidence level is insufficient, taking into account both rapid detection in normal scenarios and accurate identification in extreme scenarios.

[0035] (2) Strong environmental adaptability and sustainable model evolution. Image preprocessing parameters and edge-cloud weight coefficients can be dynamically adjusted according to visibility and camera icing, adapting to all levels of severe weather, and the environmental adaptability range is many times greater than that of traditional systems. The online transfer learning mechanism of edge-cloud collaboration enables the ANIL network to continuously receive prior knowledge from large models for parameter optimization, achieving model iterative upgrades without manual intervention, quickly adapting to new target categories and complex scenarios, and significantly reducing later maintenance costs.

[0036] (3) A high-priority early warning mechanism is designed for sudden situations such as emergency braking and illegal lane changes. The response time is ≤200ms, which is a significant improvement over traditional systems and can effectively reduce traffic accidents in severe weather. The system adopts a modular architecture with end-to-cloud collaboration and standardized interfaces. It can be directly connected to existing traffic cameras and control systems without large-scale hardware modifications. The hierarchical database can be flexibly expanded, and both large and small models can be upgraded independently. It has strong scalability and is easy to implement and promote in the field of road traffic.

[0037] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is an application architecture interaction diagram of the large-model collaborative ANIL target recognition system for camera icing and fogging provided in an embodiment of the present invention; Figure 2 The overall system workflow diagram provided for embodiments of the present invention. Detailed Implementation

[0039] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0040] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0041] like Figure 1 As shown, this embodiment provides a large-model collaborative ANIL target recognition system for cameras experiencing icing and fogging. The system adopts an edge-cloud collaborative architecture, specifically including: a data acquisition module, a wireless transmission / reception module, an edge computing module, a cloud service module, a hierarchical data storage module, and an early warning module.

[0042] The data acquisition module includes a road camera acquisition unit, multi-dimensional environmental sensors (such as temperature and humidity sensors, visibility sensors, and road icing sensors), and an on-board terminal data receiving unit. This module is responsible for real-time acquisition of traffic images, environmental data, and the status data of passing vehicles. In one optional embodiment, the environmental sensors may further include light intensity sensors, wind speed sensors, etc., to provide more dimensions of environmental characteristics. In another variation, the on-board terminal data receiving unit can communicate with the vehicle via DSRC or LTE-V2X protocols.

[0043] The edge computing module is deployed at edge nodes on the roadside or near cameras. It includes a microprocessor, an ANIL network inference unit, local temporary memory, and an image preprocessing unit. This module is responsible for adaptive preprocessing of the acquired images, calling the ANIL network for fast object detection, and calculating detection confidence. In one variation, the edge computing module can also integrate a lightweight GPU or NPU to accelerate the ANIL network inference process. In another variation, the local temporary memory can cyclically store the detection data from the most recent 10 to 30 seconds to provide short-term data caching in the event of network interruption.

[0044] The wireless transmitter / receiver module includes an end-side wireless unit and a cloud-based wireless unit, employing the 5G communication protocol to achieve high-speed data transmission between the end and the cloud, with a transmission latency of ≤100ms. Optionally, when 5G signal coverage is insufficient, 4G LTE or fiber optic communication can be used as a backup link.

[0045] The cloud service module includes the Grounding DINO 2.0 large model inference center, cloud servers, and online transfer learning units. This module is responsible for accurate object detection of low-confidence images uploaded from the client side and for optimizing the parameters of the ANIL network online using the detection results. In one variation, multiple large model instances can be deployed in the cloud to support parallel processing in different regions and traffic scenarios.

[0046] The data tiered storage module is deployed in the cloud and includes a geographic information layer database, an environmental information layer database, a traffic information layer database, a vehicle self-detection information layer database, and a management information layer database. These databases are interconnected, enabling efficient retrieval and scheduling of multi-dimensional data and providing auxiliary information for large-scale model inference in the cloud. In one variation, each database can employ a database sharding strategy, such as horizontal partitioning by time (monthly / quarterly) or by geographic region (road segment marker range), to improve retrieval efficiency under massive amounts of data.

[0047] The early warning module includes a traffic control terminal, a surrounding vehicle early warning terminal, and a road information broadcasting terminal. This module is used to display detection results and send early warning signals to relevant terminals when an emergency is detected. Optionally, early warning information can be pushed to drivers through various means such as roadside variable message signs, in-vehicle navigation systems, or mobile apps.

[0048] In addition, such as Figure 1 As shown, the system also includes a "real-time feedback" link and a "optimal model parameter" storage mechanism. The execution status of the output warning module, such as whether the warning has been confirmed by the driver, whether the control terminal has responded, and the accuracy feedback of the detection results, is transmitted back to the cloud service module in real time via a wireless transmission / reception module. This is used to further optimize the sample weights and confidence threshold strategies of the online transfer learning. After completing the ANIL parameter optimization, the cloud-based online transfer learning unit stores the updated weights and biases as optimal model parameters, associated with the current weather level and road segment features, in the management information layer of the hierarchical database, facilitating rapid retrieval or rollback in subsequent similar scenarios.

[0049] The following is combined with Figure 2 The system workflow of this invention embodiment will be described in detail.

[0050] Step S1: Multi-source data acquisition The road camera acquisition unit captures real-time traffic images at a frame rate of 30fps, while environmental sensors simultaneously collect environmental data such as visibility, camera icing level, temperature, and humidity. The vehicle-mounted terminal data receiving unit collects data on the speed, braking status, and driving trajectory of passing vehicles. All data is transmitted to the edge computing module in real time.

[0051] In one variation, the camera can capture data at a variable frame rate of 25fps to 60fps, automatically increasing the frame rate when a sharp drop in visibility is detected. In another variation, the onboard terminal data can include the vehicle's turn signal status, ABS activation status, etc., to provide richer information about vehicle dynamics.

[0052] Step S2: Adaptive preprocessing of severe weather images The image preprocessing unit performs dark channel prior dehazing and gamma correction icing enhancement on the original image based on the visibility values ​​collected by the environmental sensor and the icing level of the camera. The image preprocessing described in this step is... Figure 2 Corresponding to the image noise preprocessing in this study, its core is to dynamically adjust the defogging coefficient and gamma coefficient based on visibility and icing level to eliminate image noise and feature blur introduced by severe weather. The preprocessing parameters are dynamically adjusted according to the severity of the weather: the defogging coefficient ω=0.95 when visibility <200m; ω=0.85 when visibility ≤500m; and ω=0.75 when visibility >500m. The gamma coefficient γ=2.0 when icing level >80%; γ=1.5 when icing level ≤80%; and γ=1.2 when icing level <50%. The processed enhanced image is then transmitted to the ANIL network inference unit.

[0053] In one variation, the dehazing transmittance can be fine-tuned by incorporating temperature and humidity data: when the humidity is greater than 90%, the ω value can be appropriately reduced to avoid excessive dehazing leading to image distortion. In another variation, for mixed scenes with both icing and fogging, dehazing and icing enhancement can be performed sequentially, with the order interchangeable and determined in advance through experiments based on actual results.

[0054] Step S3: Fast inference using edge-side ANIL network The ANIL network inference unit utilizes optimized parameters (including convolutional kernel weights W and biases b) previously distributed from the cloud to perform lightweight feature extraction and object detection on the enhanced image. The feature extraction process can be represented as follows:

[0055] Where X is the preprocessed enhanced image. For depthwise separable convolution, the Swish activation function is defined as Swish(x) = x·σ(βx), with β = 1. The network output includes the target category, the target bounding box, and the edge detection confidence score. The detection results are transmitted to the microprocessor, while nearly 10 seconds of edge detection data are stored in local temporary memory.

[0056] In one variation, the ANIL network can output multiple candidate boxes and their confidence scores, not just the result with the highest confidence score. In another variation, the storage duration of the local temporary memory can be dynamically adjusted based on the storage capacity of the edge computing module, for example, storing data from the most recent 5 to 30 seconds.

[0057] Step S4: Dynamically determine confidence level based on visibility The microprocessor determines the edge-cloud weighting coefficient based on the visibility value V collected by environmental sensors. , The dynamic threshold T is calculated and determined based on a preset visibility-threshold mapping relationship. The specific calculation formula is as follows:

[0058] End-side weighting coefficient With cloud weight coefficient The calculation formula is:

[0059] Where V is the visibility value collected by the environmental sensor, in meters.

[0060] According to the above formula, when visibility > 500m (light fog / no ice), take w1 = 0.8, w2 = 0.2, T = 0.85; when 200m ≤ visibility ≤ 500m (moderate fog / light icing), take w1 = 0.6, w2 = 0.4, T = 0.90; when visibility < 200m (heavy fog / severe icing), take w1 = 0.4, w2 = 0.6, T = 0.95. If If so, the end-side detection result is directly transmitted to the output warning module; if The original image, enhanced image, and preliminary detection results from the edge are then uploaded to the cloud service module via the edge wireless unit.

[0061] In one variation, the threshold T can be further adjusted based on the camera's icing level: for example, when the icing level is greater than 80%, the threshold is increased by 0.02 to 0.05 to more strictly trigger cloud inference. In another variation, when visibility changes rapidly, such as more than 50 meters per second, the trigger threshold can be temporarily lowered to avoid a large number of missed detections due to sudden environmental changes.

[0062] Step S5: Cloud-based Grounding DINO 2.0 Precise Inference Grounding DINO 2.0's large-scale model inference center performs cross-modal semantic understanding and accurate target localization on uploaded data. Its cross-modal attention interaction can be represented as:

[0063] in, Visual features extracted for the visual Transformer For text query features, such as "vehicles", "pedestrians", "non-motorized vehicles", etc. Text query features The transpose of the matrix, The ViT visual features are weighted and mapped, where W represents the weights and T represents the visual transformer. A scaling factor is used to avoid attention value decay caused by excessively high feature dimensions. During cloud inference, multi-dimensional auxiliary data from a hierarchical data storage module is also incorporated, such as road slope, historical accident black spots, and current traffic density, to output accurate target categories, refined bounding boxes, and cloud detection confidence scores. And calculate the confidence level of edge-cloud fusion. Then, the accurate detection results are transmitted to the output early warning module.

[0064] In this embodiment, the Grounding DINO 2.0 large model employs a visual backbone network based on the Swin Transformer. Specifically, the input image is first segmented into fixed-size image patches by the Patch Partition module, and then each image patch is mapped to a feature vector through Linear Embedding. Multiple Swin Transformer Blocks sequentially perform hierarchical local-global feature extraction, with each Block containing a multi-head self-attention mechanism based on a shift window and an MLP layer. Subsequently, multi-scale features are enhanced through spatial attention and channel attention modules to highlight regions related to traffic targets. In the cross-modal fusion stage, visual features and text query features interact through a semantic-visual joint encoding module, and then pass through an IoU-aware bounding box calibration front-end filter to output high-precision target boxes and categories.

[0065] In one variation, the large cloud-based model can also perform multiple tasks simultaneously, such as vehicle type segmentation (sedan, SUV, truck, bus) and color recognition for detected vehicles. In another variation, if the uploaded image quality is still poor, such as severe icing obscuring the lens, the cloud can call upon an additional image restoration network for secondary enhancement before sending it to Grounding DINO 2.0.

[0066] The detection results output by the cloud-based large model and the output of the edge-side ANIL network are weighted and fused according to the visibility level in the dynamic weight fusion module, and the final output is the fusion confidence score. .

[0067] Step S6: Online transfer learning with edge-cloud collaboration The online transfer learning unit uses the refined bounding boxes and semantic labels output by Grounding DINO 2.0 as prior knowledge to optimize the parameters of the ANIL network. The optimization objective is to minimize the loss function of the ANIL network:

[0068] Where the weight coefficient λ = 0.5, the classification loss is... Cross-entropy loss, regression loss To smooth out the L1 loss, the convolutional kernel weights W and bias b of the ANIL network are updated after optimization, and the optimized parameters are sent to the edge ANIL network inference unit through the cloud wireless unit to achieve continuous iterative optimization of the ANIL network.

[0069] (a) Classification loss Employing cross-entropy loss, this method is designed for multi-class classification tasks involving traffic targets (vehicles, pedestrians, and non-motorized vehicles).

[0070] Where K is the number of target categories, The true category label (0-1). To predict class probabilities.

[0071] (b) Regression loss We employ smoothed L1 loss to reduce the impact of outliers on bounding box regression.

[0072] Where N is the number of bounding boxes. The coordinates of the actual target bounding box. To predict the bounding box coordinates:

[0073] In one variation, the cloud can cache optimized samples over a period of time and perform batch optimization at fixed intervals, rather than optimizing every frame, to reduce computational overhead. In another variation, online transfer learning can employ knowledge distillation, using the soft labels output by the large model as teacher signals to train the ANIL network. In yet another variation, edge computing modules in different regions can share the optimization results of the same large cloud model, but the model is fine-tuned based on the weather characteristics of each region before deployment.

[0074] The optimized ANIL network parameters are sent from the cloud wireless unit to the edge ANIL network inference unit and immediately applied to the rapid inference of the next frame image, forming a continuous cyclic optimization closed loop of "low confidence on the edge triggers fine judgment on the cloud → fine judgment results on the cloud optimize the edge model → the optimized model improves the confidence of subsequent edge inference".

[0075] Step S7: Multi-dimensional data hierarchical storage The cloud server stores the collected environmental data, traffic detection data, vehicle data, and edge-cloud detection results into corresponding databases within a hierarchical data storage module: the geographic information layer stores road slope, curve curvature, and station numbers; the environmental information layer stores visibility, icing level, temperature, humidity, and wind speed; the traffic information layer stores traffic flow, average vehicle speed, and congestion index; the vehicle detection information layer stores individual vehicle speed, braking status, and driving trajectory; and the management information layer stores device status, maintenance records, and user permissions. This interconnected storage enables rapid data retrieval and efficient scheduling, providing multi-dimensional auxiliary information for subsequent detection.

[0076] In a variation, databases can be linked through spatiotemporal indexes, such as GIS-based grid indexes, enabling queries for "all relevant data near a certain geographical location within a certain time interval" to be completed in milliseconds.

[0077] Step S8: Priority Warning for Emergencies If the edge cloud detection result determines that the vehicle is braking suddenly or changing lanes illegally, the cloud server marks the situation as high priority and outputs a warning module to send warning signals to the traffic control terminal, the surrounding vehicle warning terminal, and the road information broadcasting terminal at the same time, achieving millisecond-level response and a warning response time of ≤200ms.

[0078] In one variation, the warning signal can be set to different levels based on the severity of the emergency: Level 1 (minor) only alerts via roadside information boards; Level 2 (moderate) simultaneously sends a voice warning to vehicles behind; Level 3 (severe) automatically triggers an upstream variable speed limit sign to lower the speed limit and notifies road administration and traffic police departments. In another variation, for multiple consecutive emergencies, the system can merge them according to time sequence and spatial location to avoid repeatedly sending the same warning information to the same vehicle.

[0079] For routine test results that do not reach the level of an emergency warning, the output warning module can preview the routine test results through the traffic control terminal without triggering an emergency broadcast.

[0080] The implementation details of the core algorithm are as follows: 1. Implementation of Adaptive Image Preprocessing Algorithm (1) Dark channel prior dehazing: For the input hazy image I(x), first calculate the dark channel image and estimate the global atmospheric light value A by counting the top 0.1% of the brightest pixels. Then, select the dehazing coefficient ω based on the current visibility and substitute it into the transmittance calculation formula:

[0081] Where c represents the RGB color channel. A local window centered at x. This represents the global atmospheric light value for each channel. This represents the pixel grayscale value of any pixel y within a local window Ω(x) centered on pixel x in the original fog image, under the RGB color channel c. The value range is usually 0~255 (8-bit image).

[0082] To prevent excessively bright images due to low transmittance, a lower limit is set. (Usually taken as 0.1), finally restoring a clear image:

[0083] in, To obtain a clear image after dehazing, This is the original fog map. Transmittance, characterizing the concentration of fog. This represents the global atmospheric light value.

[0084] In practical deployment, the ω values ​​corresponding to different visibility ranges can be pre-calculated and stored in a lookup table, which the preprocessing unit can directly read without real-time calculation. In another variation, the defogging coefficient ω can be calculated based on the visibility V (in meters) using a piecewise linear function: when V < 200, ω = 0.95; when 200 ≤ V ≤ 500, ω = 0.95 - 0.1 × (V - 200) / 300; when V > 500, ω = 0.75.

[0085] (2) Gamma-corrected icing enhancement: For the input icing image (x), select the gamma correction coefficient γ based on the degree of road icing, and perform pixel-wise exponentiation: Since exponentiation can be slow on embedded devices, lookup tables or piecewise linear approximations can be used to speed it up. In a variation, the image can be converted to HSV space first, and gamma correction can be performed only on the luminance component V, while chroma and saturation remain unchanged to avoid color distortion.

[0086] 2. Implementation of ANIL Networks The ANIL network employs depthwise separable convolutions to construct a lightweight feature extractor, with the Swish activation function. The loss function is a weighted sum of cross-entropy and smoothed L1 loss. When deployed on edge, inference acceleration can be achieved using TensorFlow Lite or ONNXRuntime. In one variation, the ANIL network can further crop the number of channels or quantize the weights to reduce computational cost and memory usage. In another variation, the input image size of the ANIL network can be dynamically adjusted based on the size of the detected target: when detecting small, distant targets, the region of interest is cropped and enlarged before being fed into the network.

[0087] 3. Implementation of the Grounding DINO 2.0 large model The Grounding DINO 2.0 large-scale model is deployed in the cloud and can run on high-performance GPU servers such as NVIDIA A100 or Huawei Ascend 910. The input consists of the original and enhanced images uploaded from the client, along with text prompts (e.g., "cars, pedestrians, non-motorized vehicles, traffic signs," etc.). The model outputs bounding boxes and semantic categories. In one variation, Grounding DINO 2.0 can be fine-tuned for traffic scenarios by using a labeled dataset of icy and foggy traffic images for additional training to improve accuracy in that specific domain.

[0088] The Grounding DINO 2.0 used in this embodiment is a publicly available pre-trained model. The key is to apply it in conjunction with the ANIL lightweight network to target recognition scenarios where the camera is icing and fogging. A complete end-to-cloud collaborative system is formed through dynamic weight fusion, adaptive preprocessing, and online transfer learning.

[0089] 4. Implementation of the Dynamic Weighted Fusion Confidence Algorithm The microprocessor reads the visibility sensor values ​​in real time and looks up the values ​​in the table according to the above rules. , ,T). Calculation And compare it with T. This process can be implemented in the edge computing module through a simple state machine. In one variation, when visibility fluctuates around a threshold, such as around 200 meters, a hysteresis comparator can be introduced: the weights and T are only updated when the visibility exceeds the threshold for three consecutive samples, avoiding system jitter caused by frequent switching. In another variation, the confidence score is fused. It can also be used for subsequent early warning decisions, for example, only A situation is only considered a high-priority emergency if the value is greater than 0.9.

[0090] Those skilled in the art will understand that the above system can be implemented using the following hardware configurations, but is not limited thereto: the edge computing module can use NVIDIA Jetson Orin or Huawei Atlas 200 as the core computing unit, running the Ubuntu operating system and AI inference framework; the cloud service module can be deployed on Alibaba Cloud ECS gn7i series GPU instances, equipped with NVIDIA A10 Tensor Core GPUs; the wireless communication module uses industrial-grade 5G modules, supporting SA / NSA networking; and the environmental sensors can use the Vaisala PTU310 integrated sensor and the Lufft IRS31 Pro road condition sensor.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A large-scale collaborative ANIL target recognition method for camera icing and fogging, characterized in that, Includes the following steps: S1: Multi-source data acquisition, acquiring traffic images captured by road cameras, visibility and camera icing levels captured by environmental sensors, and vehicle status data captured by vehicle terminals; S2: Based on the visibility and camera icing, dynamically adjust the dark channel prior dehazing coefficient ω and gamma correction coefficient γ, and perform adaptive preprocessing on the original image to obtain an enhanced image; The prior defogging coefficient ω for the dark channel is dynamically adjusted according to visibility: ω = 0.95 when visibility < 200m; ω = 0.85 when visibility ≤ 500m; and ω = 0.75 when visibility > 500m. The gamma correction coefficient γ is dynamically adjusted according to the camera's icing level: γ = 2.0 when icing level > 80%; γ = 1.5 when icing level ≤ 80%; and γ = 1.2 when icing level < 50%. S3: Perform target detection on the enhanced image using an edge-side ANIL network, and output the target category, target bounding box, and edge-side detection confidence score. The edge-side ANIL network employs depthwise separable convolutions to construct a lightweight feature extractor. S4: Determine the dynamic threshold T according to the visibility. If ≥T, directly output the end - side detection result to the warning module; if <T, upload the original image, the enhanced image and the end - side detection result to the cloud; where the lower the visibility, the higher the dynamic threshold T; S5: Utilize the cloud-based Grounding DINO 2.0 large-scale model to perform cross-modal semantic understanding and accurate target localization on the uploaded data. Combined with the vehicle status data, determine the vehicle behavior state and output accurate target category, refined localization box, and cloud-based detection confidence score. ; S6: Using the refined localization boxes and semantic labels output by Grounding DINO 2.0 as prior knowledge, perform online transfer learning on the edge ANIL network to optimize the convolutional kernel weights W and bias b of the ANIL network, and then send the optimized parameters down to the edge.

2. The method according to claim 1, characterized in that, In S2, the formula for calculating the transmittance t(x) of the dark channel prior dehazing is: Where c is the RGB color channel, and Ω(x) is the local window centered at x. This represents the global atmospheric light value for each channel. This represents the pixel grayscale value of any pixel y within a local window Ω(x) centered on pixel x in the original fog image, under the RGB color channel c. Restore clear images by: in, To obtain a clear image after dehazing, This is the lower limit of transmittance. This is the original fog map. Transmittance, This represents the global atmospheric light value.

3. The method according to claim 1, characterized in that, In S4, the dynamic threshold T is determined based on visibility: when visibility > 500m, T = 0.85; when 200m ≤ visibility ≤ 500m, T = 0.90; when visibility < 200m, T = 0.

95.

4. The method according to claim 1, characterized in that, S5 also includes: determining end-side weights based on visibility. and cloud weight And calculate the confidence level of edge-cloud fusion. Among them, when visibility is >500m, =0.8, =0.2; when 200m ≤ visibility ≤ 500m, =0.6, =0.4; when visibility <200m, =0.4, =0.

6.

5. The method according to claim 1, characterized in that, In S6, the optimization objective of the online transfer learning is to minimize the loss function of the ANIL network. Where the weight coefficient λ=0.5, and the classification loss Cross-entropy loss, regression loss To smooth out the L1 loss.

6. The method according to claim 1, characterized in that, It also includes S7: storing the collected environmental data, traffic detection data, vehicle data, and edge-cloud detection results into the corresponding databases of the layered data storage module, wherein the layered data storage module includes a geographic information layer database, an environmental information layer database, a traffic information layer database, a vehicle self-detection information layer database, and a management information layer database.

7. The method according to claim 1, characterized in that, It also includes S8: When the end-cloud detection result determines that the vehicle is braking suddenly or changing lanes illegally, the sudden situation is marked as high priority, and the warning signal is sent to the traffic control terminal, the surrounding vehicle warning terminal and the road information broadcasting terminal simultaneously through the output warning module, with a warning response time of ≤200ms.

8. The method according to claim 1, characterized in that, The edge-side ANIL network employs depthwise separable convolutions to construct a lightweight feature extractor. Its feature extraction is as follows: Where X is the preprocessed enhanced image. For depthwise separable convolution, the Swish activation function is defined as Swish(x) = x·σ(βx), with β = 1.

9. The method according to claim 1, characterized in that, The cross-modal attention interaction of the Grounding DINO 2.0 large model is as follows: in, Visual features extracted for the visual Transformer For text query features, Text query features The transpose of the matrix, The ViT visual features are obtained after weight mapping, where W represents the weights and T represents the visual transformer. This is the scaling factor.

10. A large-model cooperative ANIL target recognition system for camera icing and fogging using the method described in any one of claims 1-9, characterized in that, The architecture adopts an edge-cloud collaborative approach, including: The data acquisition module includes a road camera acquisition unit, a multi-dimensional environmental sensor, and an on-board terminal data receiving unit, used to acquire traffic images acquired by the road camera, visibility and camera icing level acquired by the environmental sensor, and vehicle status data acquired by the on-board terminal. The edge computing module, deployed on the roadside, includes a microprocessor, an ANIL network inference unit, a local temporary memory, and an image preprocessing unit. It is used to adaptively preprocess the original image to obtain an enhanced image based on the visibility and the icing degree of the camera, perform target detection on the enhanced image using the ANIL network, and determine a dynamic threshold T based on the visibility to determine whether to trigger cloud inference. The wireless transmitter / receiver module uses the 5G communication protocol to achieve high-speed data transmission between the device and the cloud. The cloud service module includes the Grounding DINO 2.0 large model inference center, cloud server and online transfer learning unit, which is used to perform cross-modal semantic understanding and accurate target localization on uploaded data, and combine the vehicle status data to determine the vehicle behavior status, and use the detection results as prior knowledge to perform online transfer learning on ANIL network. The data tiered storage module, deployed in the cloud, includes a geographic information layer database, an environmental information layer database, a traffic information layer database, a vehicle self-detection information layer database, and a management information layer database. The output warning module includes a traffic control terminal, a surrounding vehicle warning terminal, and a road information broadcasting terminal, which are used to output detection results and warnings of emergencies.

Images