Road construction scene identification method and device, equipment and storage medium

Abstract

The embodiment of the invention provides a road construction scene identification method and device, equipment and a storage medium. The method comprises the steps of obtaining a road image of a road where a vehicle is located at the current moment; respectively inputting the road image into at least two preset detection models to obtain detection results output by the preset detection models; wherein the preset detection model is a pre-trained neural network model, the preset detection model is used for identifying a construction object of a preset category, and the detection result represents a distribution condition of the construction object of the preset category in the road image; according to detection results output by the preset detection models, construction scene information of a road where the vehicle is located is determined; wherein the construction scene information represents the current construction state of the road. According to different types of construction objects, different detection models are set for targeted detection, and the recognition precision of the construction scene is improved in combination with the detection results of the multiple detection models.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment and storage medium for recognizing road construction scenarios. Background Technology

[0002] With the continuous advancement of transportation infrastructure construction, road construction is widely present in various projects such as highway reconstruction and expansion, urban arterial road maintenance, and bridge and tunnel construction. Construction areas not only occupy road resources but also alter the original traffic flow, directly impacting traffic safety and efficiency.

[0003] Therefore, it is necessary to conduct timely and accurate inspections of road construction conditions to improve driving safety and the driving experience. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for identifying road construction scenes, in order to improve the accuracy of construction scene identification.

[0005] In a first aspect, embodiments of this application provide a method for identifying road construction scenarios, including:

[0006] Obtain a road image of the road where the vehicle is located at the current moment;

[0007] The road image is input into at least two preset detection models to obtain the detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image;

[0008] Based on the detection results output by each of the preset detection models, the construction scene information of the road where the vehicle is located is determined; wherein, the construction scene information represents the current construction status of the road.

[0009] Secondly, embodiments of this application provide a road construction scene identification device, comprising:

[0010] The image acquisition unit is used to acquire road images of the road where the vehicle is located at the current moment.

[0011] An image detection unit is used to input the road image into at least two preset detection models to obtain the detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image;

[0012] The scene determination unit is used to determine the construction scene information of the road where the vehicle is located based on the detection results output by each of the preset detection models; wherein the construction scene information represents the current construction state of the road.

[0013] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0014] The memory stores computer-executed instructions;

[0015] The processor executes computer execution instructions stored in the memory, causing the processor to perform the implementation method described in the first aspect above.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the embodiments of the first aspect above.

[0017] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the implementation methods described in the first aspect above.

[0018] This application provides a method, apparatus, device, and storage medium for identifying road construction scenes. It acquires real-time road images of the road where the vehicle is currently located and inputs these images into different preset detection models. Each preset detection model detects a preset category of construction objects. Employing at least two preset detection models in parallel to identify different categories of construction objects improves the detection accuracy for different categories. By fusing the detection results of multiple preset detection models, the overall construction scene information is determined, solving problems such as weak generalization ability of single models and susceptibility to interference from single features. This effectively improves the detection accuracy of road construction scenes, thereby enhancing vehicle driving safety and the user's driving experience. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 A flowchart illustrating a method for identifying road construction scenarios provided in an embodiment of this application;

[0021] Figure 2 A flowchart illustrating a method for identifying road construction scenarios provided in an embodiment of this application;

[0022] Figure 3 This is a schematic diagram illustrating the reasoning process of the preset detection model provided in the embodiments of this application;

[0023] Figure 4 A flowchart illustrating a method for identifying road construction scenarios provided in an embodiment of this application;

[0024] Figure 5 A schematic diagram of the structure of a road construction scene recognition device provided in an embodiment of this application;

[0025] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0026] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0028] First, let me explain the terms used in this application:

[0029] Water-filled barriers: a type of construction barrier commonly used in construction sites, road repairs, temporary enclosures, and other scenarios, serving as a traffic control function to guide traffic flow and change driving routes;

[0030] Construction barriers: Objects or facilities set up on roads or specific areas to prevent, restrict, guide or warn vehicles or pedestrians from passing through. Their core purpose is to manage traffic flow, ensure safety, and delineate restricted or dangerous areas. For example, traffic cones.

[0031] Dashcam: A vehicle-mounted device used for real-time monitoring and dynamic early warning of road construction scenarios.

[0032] With the continuous advancement of transportation infrastructure construction, the road network is constantly being improved. Construction scenarios are widespread during road construction, maintenance, and repair, encompassing various types such as urban arterial road reconstruction and expansion, highway repair, and highway reconstruction and expansion. Construction areas not only occupy road resources but also alter existing traffic flow, directly impacting traffic safety and efficiency. Identifying road construction scenarios is a crucial link connecting construction management and traffic control, directly determining the effectiveness of construction safety, traffic safety, traffic management, and compliance supervision.

[0033] Current construction scene recognition technologies have significant limitations: manual inspections rely on subjective judgment and cannot cover all weather conditions and all road sections; fixed surveillance cameras are costly to deploy, have limited viewing angles, and are severely affected by rain, fog, and lighting conditions; anomaly inference methods based on GPS trajectories are prone to misclassifying non-construction factors such as traffic accidents, temporary traffic control, and signal drift as construction events, leading to high false alarm rates and delayed responses. Especially on resource-constrained devices such as dashcams, traditional general-purpose target detection models, due to their large number of parameters and computational intensity, struggle to meet the combined constraints of real-time performance, limited memory, and lightweight NPU (Neural Processing Unit) computing power, resulting in low accuracy in construction element recognition, large inter-frame jitter, and unstable state determination. Therefore, there is an urgent need for a road construction scene recognition method that is adaptable to edge deployment and balances robustness in recognizing multiple types of construction objects with inference efficiency.

[0034] This application provides a method, apparatus, equipment, and storage medium for identifying road construction scenarios, aiming to solve the above-mentioned technical problems in the prior art.

[0035] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0036] Figure 1 This is a flowchart illustrating a method for identifying road construction scenes according to an embodiment of this application. This method can be executed by a road construction scene identification device. Figure 1 As shown, the method includes:

[0037] S101. Obtain the road image of the road where the vehicle is located at the current time.

[0038] For example, the vehicle may be equipped with an image acquisition device, such as a dashcam, to acquire video streams of the vehicle's surroundings in real time or at regular intervals, wherein the image frames in the video stream are road images of the road where the vehicle is located.

[0039] The current moment refers to a specific sampling point in the continuous frame sequence of the dashcam's video stream, and its timestamp is synchronized with the system's real-time clock. The road image of the road where the vehicle is located is a single-frame image captured by the dashcam's camera, with an image resolution of 1280×720 or 1920×1080. The current road image can be acquired in real time while the vehicle is in motion. Road images from different directions can be obtained simultaneously, and these images can also be stitched together to obtain a complete road image.

[0040] S102. Input the road image into at least two preset detection models respectively to obtain the detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image.

[0041] For example, at least two preset detection models are pre-built and trained. These preset detection models are lightweight, dedicated neural network models deployed on the edge. Each preset detection model can be structurally customized to meet different semantic granularity and spatial representation requirements. Each preset detection model corresponds one-to-one with a preset category; that is, the preset detection model is used to identify construction objects of the corresponding preset category. The preset category is a pre-defined category of construction objects, which can characterize the attributes, types, appearance, and other features of the construction objects.

[0042] Each preset detection model can perform inference in parallel. That is, the road image can be input into each preset detection model, and each preset detection model performs image processing and outputs its own detection results. The detection results can characterize the distribution of construction objects of preset categories in the road image. For example, it can characterize whether construction objects exist in the road image, the number of construction objects, and the category of construction objects.

[0043] In this embodiment, the model architecture of each preset detection model is not specifically limited. For example, two types of preset detection models are set up. The first type is a lightweight object detection model, used to identify discrete construction objects with clear bounding boxes, including traffic signs (such as reflective signs such as "Construction Ahead," "Slow Down," and "Road Closed") and construction barriers (such as traffic cones and crash barriers). The second type is a lightweight semantic segmentation model, used to identify planar construction objects with continuous contour boundaries, typically represented by water-filled barriers. The two types of preset detection models can be integrated and deployed using the ONNX (Open Neural Network Exchange) intermediate representation format, sharing the input image normalization preprocessing pipeline to avoid repeated loading and memory copying, saving memory and NPU scheduling overhead. That is, the road image can be input into the image normalization preprocessing pipeline, preprocessed, and then sent to each preset detection model.

[0044] In this embodiment, the preset categories can be dynamically expanded according to construction specifications. For example, new categories such as "nighttime warning lights", "mobile construction signs", and "asphalt pavers" can be added. The corresponding models can be fine-tuned locally on the end side through incremental learning, without having to re-upload the full model.

[0045] The detection results can be structured data packets. For object detection models, the detection results can include the coordinates, confidence scores, and category IDs of each detection box. For semantic segmentation models, the detection results can include contour point sets, bounding rectangles, connected component areas, and mask hash values.

[0046] S103. Based on the detection results output by each preset detection model, determine the construction scene information of the road where the vehicle is located; wherein, the construction scene information represents the current construction status of the road.

[0047] For example, the detection results output by each preset detection model constitute a multi-source heterogeneous data set, and its fusion strategy can rely on weighted average or hard voting, or it can adopt a logical judgment mechanism based on a rule engine.

[0048] For example, first, perform single-frame validity filtering on each type of detection result: for the object detection model, non-maximum suppression can be applied to eliminate overlapping boxes, and confidence threshold and minimum size threshold can be set; for the semantic segmentation model, closed contours are extracted, only valid contours with an area greater than 100 pixels are retained, and noise spots are removed; then, cross-model logical association is performed: if more than 4 traffic cones are detected and the total area of ​​the water-filled guardrail contours is greater than 500 pixels, then the construction scene information is determined to be that there is construction on the current road.

[0049] By acquiring road images in real time from the vehicle, at least two lightweight neural network models with heterogeneous structures and task specializations are used to extract the spatial distribution features of construction objects of different shapes in parallel. Then, based on the logical fusion of multi-source detection results, construction scene information is generated. Different preset detection models are designed for different construction objects, and the two complement each other to avoid the blind spots of single models under occlusion, scale changes, and lighting interference. The lightweight model design and ONNX integration ensure real-time inference capabilities on low-computing-power devices such as dashcams. This embodiment effectively solves the problems of poor real-time performance, low robustness, high false positive rate, and inability to adapt to low-computing-power devices in vehicle-mounted construction recognition, achieving a comprehensive technical effect of high accuracy, low latency, strong anti-interference, and deployability.

[0050] This application provides a method for identifying road construction scenes. It acquires real-time road images of the road where the vehicle is currently located and inputs these images into different preset detection models. Each preset detection model detects a specific category of construction objects. Employing at least two preset detection models in parallel to identify different categories of construction objects improves the detection accuracy for different categories. By fusing the detection results from multiple preset detection models, the overall construction scene information is determined, solving problems such as weak generalization ability of single models and susceptibility to interference from single features. This effectively improves the detection accuracy of road construction scenes, thereby enhancing vehicle driving safety and the user's driving experience.

[0051] Figure 2 A flowchart illustrating a method for identifying road construction scenes provided in this application embodiment is shown below. Figure 2 As shown, this embodiment, based on the above embodiments, provides a detailed description of a method for identifying road construction scenarios. The method includes:

[0052] S201. Obtain the road image of the road where the vehicle is located at the current time.

[0053] For example, this step can refer to step S101 above, and will not be repeated here.

[0054] S202. Based on the road image and the first detection model, a first detection result is obtained; wherein, the first detection result is the detection result output by the first detection model, and the first detection result represents the distribution of road administration facilities in the road image.

[0055] For example, the preset detection model refers to a lightweight neural network model deployed on an in-vehicle device (such as a dashcam), trained offline and quantized for compression. It is used for real-time edge-side recognition of specific construction-related objects in road images under resource-constrained conditions, supporting synchronous or asynchronous inference for single-frame images. The first detection model is specifically designed to identify road infrastructure, with preset categories covering discrete, small-scale, high-contrast targets such as traffic signs and construction barriers. The second detection model is specifically designed to identify construction guardrails, with preset categories specifically referring to continuous, regionalized, low-texture but with significant contour and area features, such as water-filled barriers, movable barriers, and corrugated beam temporary barriers. Different preset detection models can differ in their model architecture. In this embodiment, two preset detection models can be used: a first detection model and a second detection model. The first detection model adopts an object detection paradigm, using YOLOv8-Nano as the backbone network, CSPPAN (CrossStage Partial Path Aggregation Network) as the network neck, and NanoDetPlusHead as the network detection head, outputting bounding box coordinates and category labels with confidence. The second detection model adopts a semantic segmentation paradigm, using PIDNet (Pixel Intermediate Derivative Network) as the backbone network, and FCN-SegHead (Fully Convolutional Network SegmentationHead) as the network head, outputting pixel-level classification masks, and then extracting connected components and calculating spatial proportions through post-processing.

[0056] The dual-model design in this embodiment is not a simple juxtaposition, but rather a task decoupling based on prior knowledge of the physical form and visual representation of the construction object. Road infrastructure is often distributed as isolated points or small rectangles, emphasizing precise positioning and fine-grained classification capabilities; construction guardrails, on the other hand, present long strips, continuous bands, or block-like structures, emphasizing contour integrity and area robustness. The two models are differentiated and adapted in terms of receptive field scale, feature downsampling depth, and loss function weight configuration. For example, the first detection model can be replaced with EfficientDet-Lite3 or PP-YOLOE-Lite, and the second detection model can be replaced with BiSeNetV2 or SegFormer-B0.

[0057] The road image is a RGB (Red, Green, Blue) frame image captured by a dashcam at 15 FPS (Frames Per Second), with a selectable resolution of 1280×720. The first detection result is structured data output by the first detection model after performing forward inference on the input image, containing at least one detection instance, i.e., a construction object. Each instance corresponds to the position of an outer rectangle, a confidence score, and a category ID. The confidence score represents the probability that the instance within the rectangle is a road maintenance facility, and the category ID corresponds to categories such as traffic signs or construction barriers. The first detection result directly reflects the two-dimensional spatial distribution density, relative positional relationship, and category composition of road maintenance facilities in the image space. For example, three traffic cones were detected in the lower left quadrant of the image, and one "construction area" was detected in the upper right quadrant.

[0058] In this embodiment, road administration facilities include construction roadblocks; the first detection model includes at least two downsampling convolutional layers; based on the road image and the first detection model, a first detection result is obtained, including: inputting the road image into the downsampling convolutional layer in the first detection model to obtain the output features corresponding to the downsampling convolutional layer; wherein, the output features are feature vectors representing the road image; based on the output features corresponding to each downsampling convolutional layer, the construction roadblocks in the road image are selected to obtain the number of construction roadblocks; if the number of construction roadblocks is greater than a preset number threshold, the first detection result is determined to be the presence of road administration facilities in the road image, and the first detection result is stored in a preset first cache pool.

[0059] Specifically, road administration facilities are physical devices used for traffic guidance, area isolation, and safety warnings in road management and maintenance operations. Their scope includes construction barriers, traffic signs, traffic cones, water-filled barriers, crash barriers, temporary traffic lights, etc. In this embodiment, road administration facilities may include construction barriers. Construction barriers can be typical low-lying obstacles that are small in size, regularly shaped, highly reflective, and easily affected by light and obstruction; for example, they can be traffic cones, commonly known as cones. In the RGB images captured by a dashcam, construction barriers typically appear as high-contrast color blocks with a bright top and a dark bottom, with pixel sizes usually concentrated in the range of 12×12 to 32×32.

[0060] Downsampling convolutional layers are neural network layers that achieve spatial dimension compression. Their core function is to progressively reduce the resolution of feature maps, expand the receptive field, and enhance semantic abstraction capabilities. The first detection model can include multiple downsampling convolutional layers, each with a different stride, and these layers can be connected in parallel. Considering the small size of construction roadblocks and their limited pixel range in road images, only two strides (8 and 16) were set in the design of the construction roadblock detection head. This avoids problems such as insufficient information about construction roadblock elements in the feature map due to excessive downsampling sizes. A downsampling convolutional layer with a stride of 8 can be used to capture details, while a downsampling convolutional layer with a stride of 16 can be used to capture contours and contextual information.

[0061] After the road image enters the first detection model, it can pass through various downsampling convolutional layers. Each downsampling convolutional layer performs downsampling processing and outputs its corresponding output features, which represent the downsampling road image. For example, there may be a first downsampling convolutional layer and a second downsampling convolutional layer, each outputting different features.

[0062] After obtaining the output features corresponding to each downsampled convolutional layer, the output features can be fused. Based on the fused feature vector, construction obstacles in the road image can be bounded to obtain the number of construction obstacles. For example, a feature pyramid fusion strategy can be used. The output features of the first downsampled convolutional layer and the output features of the second downsampled convolutional layer are upsampled and added element-wise to obtain a fused feature vector, which is then fed into a lightweight detection head. The detection head is a NanoDetPlusHead structure with two parallel branches. The localization branch outputs bounding box regression parameters, and the classification branch outputs the confidence score of the construction obstacle category. The bounding box selection process includes: calculating the intersection-union score for each effective anchor point of the fused feature vector, removing overlapping boxes through non-maximum suppression, and finally retaining the rectangular boxes with a confidence score greater than 0.5, thereby obtaining the number of construction obstacles.

[0063] A pre-set quantity threshold is established; for example, the preset quantity threshold ranges from 3 to 6. This means that a single construction obstacle rarely appears in isolation in a properly deployed scenario; typically, four or more traffic cones form the smallest effective isolation unit. The number of detected construction obstacles is compared with the preset quantity threshold. If the number of construction obstacles is less than or equal to the preset quantity threshold, a false positive is considered, such as headlight reflections, road surface oil stains, or small stones, and the first detection result is that no road maintenance facilities are present in the road image. If the number of construction obstacles is greater than the preset quantity threshold, the first detection result is that road maintenance facilities are present in the road image.

[0064] A first buffer pool is pre-configured. This first buffer pool can be a pre-allocated circular buffer in memory with a capacity of 30 frames. It supports FIFO (First-In-First-Out) read and write operations, its address space is independent of the main inference thread, and it is protected by a mutex lock to ensure data consistency in a multi-threaded environment. After obtaining the first detection result, it can be stored in the first buffer pool. For example, the first detection result can be associated with road image markers or with the current time.

[0065] The beneficial effects of this setup are that the construction obstacle existence determination mechanism based on multi-scale downsampling feature fusion not only accurately captures the outline of small-sized obstacles while preserving spatial details, but also suppresses background interference through semantic context, thus improving the recognition accuracy of construction obstacles. By setting a quantity threshold, false alarms caused by single-frame false detections are avoided, further improving recognition accuracy. Storing the first detection result in a first cache pool allows for decoupling from subsequent scene recognition and other processes, enabling asynchronous execution of each process and improving recognition efficiency.

[0066] In this embodiment, road infrastructure includes traffic signs; the first detection model includes at least two downsampling convolutional layers; based on the road image and the first detection model, a first detection result is obtained, including: inputting the road image into the downsampling convolutional layer in the first detection model to obtain the output features corresponding to the downsampling convolutional layer; wherein, the output features are feature vectors representing the road image; if it is determined that the road image contains traffic signs based on the output features corresponding to each downsampling convolutional layer, then the traffic signs in the road image are classified to determine the type information of the traffic signs; wherein, the type information represents the function of the traffic signs; if it is determined that the type information of the traffic signs meets the preset conditions, then the type information of the traffic signs is determined as the first detection result, and the first detection result is stored in a preset first cache pool.

[0067] Specifically, road infrastructure can include traffic signs, which are signs placed along roadsides. These signs may be used to warn of construction or are simply standard traffic management signs. Therefore, after identifying traffic signs in a road image, it is necessary to further determine the type of traffic sign to avoid misjudging the construction scenario.

[0068] The first detection model includes at least two downsampling convolutional layers. Since traffic signs are larger than construction barriers, the number of downsampling convolutional layers and the stride can be greater for traffic signs. For example, three downsampling convolutional layers can be set with strides of 8, 16, and 32 respectively. This achieves a balance between localization accuracy for small targets (such as distant traffic signs) and global context modeling capabilities. Lower-level downsampling feature maps retain high-resolution details, supporting accurate sign border localization, while higher-level downsampling feature maps have a larger receptive field, supporting robust discrimination under complex backgrounds (such as tree shading or rain / fog).

[0069] The road image is input into the downsampling convolutional layers of the first detection model to obtain the output features corresponding to each downsampling convolutional layer. By fusing the output features, it is determined whether the road image contains traffic signs. For example, after global average pooling or a feature pyramid fusion network, the feature vector can be further compressed into a one-dimensional feature vector, which is used as the input of the subsequent detection head. The fused feature vector implicitly encodes multi-dimensional information such as texture, edges, color distribution, and spatial layout in the image, and is the essential criterion for distinguishing traffic signs from other interference objects (such as billboards, lampposts, and vehicle reflective stickers).

[0070] A preset multi-scale feature fusion strategy can be used to fuse the various output features. In this embodiment, the preset multi-scale feature fusion strategy is not specifically limited. For example, the output feature maps of different downsampling convolutional layers can be horizontally connected and upsampled to construct a feature pyramid. Based on this, the detection head independently predicts the sign candidate box and its confidence score for each feature pyramid level. If the confidence score of the candidate box is ≥0.6, the content in the candidate box is considered a valid traffic sign.

[0071] If no traffic signs are detected in the road image, the first detection result is that there are no road maintenance facilities in the road image. If traffic signs are detected in the road image, further classification processing can be performed on the traffic signs to determine their type information. Type information characterizes the function of the traffic sign; for example, it can determine whether the traffic sign is used for construction site warnings or routine traffic management. For example, the image region corresponding to the rectangular frame of each traffic sign can be extracted from the road image and input into a preset classification model to obtain the type information corresponding to the traffic sign. The classification model is a pre-trained neural network model specifically designed for classifying traffic signs. In this embodiment, the architecture of the classification model is not specifically limited.

[0072] Pre-set conditions are established, such as allowing for multiple preset sign types that represent traffic signs used in construction scenarios. If the traffic sign type information meets the preset conditions—that is, if the traffic sign type information belongs to a preset sign type—then that traffic sign type information is identified as the first detection result. In other words, the first detection result may contain the traffic sign type information. If the traffic sign type information does not meet the preset conditions, then that traffic sign will not participate in subsequent construction scenario decisions.

[0073] The number of traffic signs that meet preset conditions and the type information of the traffic signs can be used together as the first detection result, and the first detection result can be stored in a preset first cache pool. If the road image contains both construction barriers and traffic signs that meet preset conditions, the number of construction barriers, the number of traffic signs that meet preset conditions, and the type information of the traffic signs can be used together as the first detection result and stored.

[0074] The beneficial effect of this setup is that, in response to the problem that there are many types of traffic signs but only a portion of them have construction indication significance, a two-level discrimination mechanism of detection and classification is constructed. Robust features are extracted using downsampling convolutional layers, and interference from non-construction signs is eliminated through whitelist filtering. This avoids misjudging ordinary traffic management signs as construction signals, significantly improving the accuracy of sign recognition, and thus improving the recognition accuracy of construction scenarios.

[0075] In this embodiment, traffic signs in road images are classified to determine their type information. This includes: obtaining the image region of the traffic sign from the road image and scaling the image region based on a preset image range; and determining the type information of the traffic sign based on the scaled image region.

[0076] Specifically, based on the output features of each downsampling convolutional layer, a bounding box can be selected on the road image to extract the image region of the traffic sign, i.e., the image content within the rectangular box. This image region is in RGB format, and its size dynamically changes with the distance between the vehicle and the sign, the installation height, and the camera's tilt angle.

[0077] After obtaining the image region, it can be scaled based on a preset image range to uniformly adjust the cropped image region to a fixed input size. The preset image range can be 128×128 pixels. The scaling process adopts a proportional scaling strategy: first, the long side of the image region is scaled proportionally to 128, and then the short side is zero-padded to 128 pixels to maintain the original aspect ratio and structural integrity, avoiding character deformation, border distortion, or color distortion caused by stretching.

[0078] The scaled image region is obtained and input into a pre-defined classification model, which then outputs the traffic sign type information. In other words, a standardized 128×128 pixel image is input into a pre-trained traffic sign classification model, which outputs the optimal matching type information. The classification model in this embodiment uses MobileOne-S as its backbone, embeds a RepOptimizer reparameterization structure, supports INT8 quantization after training, and utilizes an NPU acceleration unit for forward propagation during inference. The output layer is a Softmax classifier, and the type information includes, but is not limited to, construction warnings, speed limits, and directional signs.

[0079] The beneficial effects of this setup are that it enables high-fidelity extraction of traffic signs from dashcam video streams, robust scaling, and fine-grained business classification, supports semantic-level accurate determination of construction scene information, and improves the generalization ability of construction element recognition and the reliability of project implementation.

[0080] S203. Based on the road image and the second detection model, a second detection result is obtained; wherein, the second detection result is the detection result output by the second detection model, and the second detection result represents the distribution of construction guardrails in the road image.

[0081] For example, the second detection result is the output of the second detection model after performing semantic segmentation inference on the same road image. For instance, the second detection model can convert the road image into a binary mask image through semantic segmentation, where a pixel value of 1 represents a region belonging to a construction guardrail, and the remaining pixels have a value of 0. The `findContours` algorithm from OpenCV can be used to extract the contour of the largest connected component in the binary mask image. Further calculations are made of the area enclosed by the contour, the contour perimeter, aspect ratio, and the total number of pixels. Finally, the "area information" is output as the core quantification indicator, which represents the distribution of construction guardrails in the road image—that is, the second detection result. The unit of area information can be the number of image pixels, representing the proportion of visual space occupied by the construction guardrail in the current field of view, rather than its physical size. For example, if a horizontal water-filled barrier is detected, its mask coverage area is 2147 pixels, corresponding to 1.3% of the effective screen area in the image. This value can be compared with a preset area threshold to determine whether there is an effective construction barrier. This avoids the problem of inaccurate selection caused by barrier deformation, bending, and partial occlusion in traditional target detection, and significantly improves the stability under complex road curvature and multi-angle shooting.

[0082] In this embodiment, the execution order between steps S202 and S203 is not specifically limited. Figure 3 This is a schematic diagram of the reasoning process of the preset detection model. Figure 3 In this process, a road image is first input, undergoes uniform preprocessing, and then inputs into two model branches: the left branch is the first detection model, and the right branch is the second detection model. The detection result output by the first detection model is the first detection result, and the detection result output by the second detection model is the second detection result.

[0083] This embodiment achieves differentiated modeling and collaborative identification of heterogeneous objects in construction scenarios. The first detection model focuses on the existence and category identification of discrete road infrastructure, providing semantic support for construction intent; the second detection model focuses on the spatial coverage analysis of continuous construction guardrails, providing structural basis for construction scale and regional enclosure. The detection results output by the two models are complementary at the logical level, avoiding false alarms caused by false detections from a single model and improving the identification accuracy of construction scenarios.

[0084] In this embodiment, based on the road image and the second detection model, a second detection result is obtained, including: inputting the road image into the second detection model to determine the area information of the construction guardrail in the road image; wherein, the area information represents the number of pixels occupied by the construction guardrail; if the area information of the construction guardrail is greater than a preset area threshold, the second detection result is determined to be that there is a construction guardrail in the road image, and the second detection result is stored in a preset second cache pool.

[0085] Specifically, the second detection model is a pre-trained lightweight semantic segmentation model. Its backbone network can be PIDNet, the head can be FCN-SegHead, and the output data can be a pixel-wise category mask with the same resolution as the input image. In this mask, only the pixel positions marked as "construction guardrail" are 1, and the rest are 0.

[0086] Construction guardrails can refer to physical isolation facilities commonly seen in road construction scenarios, such as water-filled barriers, plastic isolation piers, and continuous reflective guardrails, which have large areas of color blocks, high-contrast edges, and regular outlines. Their visual representation is usually a continuously distributed red / orange / yellow rectangular or trapezoidal area, which is usually presented in the image as a closed area with high connectivity and an aspect ratio between 1:3 and 3:1.

[0087] Area information can refer to the number of positive sample pixels obtained by counting pixels in the binary mask output by the second detection model, that is, the number of all pixels that the model classifies as "construction guardrail". For example, when a water-filled barrier is partially obscured by a truck, the mask can still accurately count the number of pixels in the visible part, while traditional YOLO bounding box detection may miss detections due to decreased confidence.

[0088] In this embodiment, the area information can also be post-processed on the original mask through morphological closing operations to fill small holes and connect broken pixels. Then, connected component analysis is performed to retain only the number of pixels in the largest connected component as the final area information, so as to further suppress noise interference.

[0089] The preset area threshold can range from 80 to 200 pixels, with 100 pixels being an optional value. The second detection result can characterize whether a construction guardrail exists in the road image. If it does, the second detection result can also include the area information of the construction guardrail. The preset second buffer pool is a fixed-size circular buffer in memory with a capacity of 32 frames, implemented using a thread-safe lock-free queue. The second detection result can be associated with the road image and stored in the preset second buffer pool.

[0090] The advantage of this setup is that it allows for robust determination of the existence of construction barriers on low-computing-power vehicle-mounted terminals using pixel counting—a computationally inexpensive and highly interference-resistant method. By setting an area threshold, false positives for construction barriers are avoided. Furthermore, storing the second detection result in a second cache pool decouples the various stages of the construction scene recognition process, improving recognition efficiency.

[0091] S204. Based on the detection results output by each preset detection model, determine the construction scene information of the road where the vehicle is located; wherein, the construction scene information represents the current construction status of the road.

[0092] For example, this step can refer to step S103 above, and will not be repeated here.

[0093] This application provides a method for identifying road construction scenes. It acquires real-time road images of the road where the vehicle is currently located and inputs these images into different preset detection models. Each preset detection model detects a preset category of construction objects. By employing at least two preset detection models in parallel to identify different categories of construction objects, the detection accuracy for different categories of construction objects is improved. The detection results of multiple preset detection models are fused to determine the overall construction scene information, solving problems such as weak generalization ability of single models and susceptibility to interference from single features. This effectively improves the detection accuracy of road construction scenes, thereby enhancing vehicle driving safety and the user's driving experience.

[0094] Figure 4 A flowchart illustrating a method for identifying road construction scenes provided in this application embodiment is shown below. Figure 4 As shown, this embodiment, based on the above embodiments, provides a detailed description of a method for identifying road construction scenarios. The method includes:

[0095] S401. Obtain the road image of the road where the vehicle is located at the current time.

[0096] For example, this step can refer to step S101 above, and will not be repeated here.

[0097] S402. Input the road image into at least two preset detection models respectively to obtain the detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image.

[0098] For example, this step can refer to step S102 above, and will not be repeated here.

[0099] S403. Based on the detection results output by each preset detection model, determine the identification information of the road where the vehicle is located at the current moment; wherein, the identification information indicates whether there are construction objects in the road image at the current moment.

[0100] For example, the detection results output by each preset detection model may include a first detection result and a second detection result, that is, it may include, but is not limited to: the existence determination result of the road administration facilities output by the first detection model and the existence determination result of the construction guardrail output by the second detection model. By combining the detection results of each preset detection model, it can be determined whether or not a construction object exists in the road image, thereby obtaining the identification information of the road image at the current moment. The identification information is used to characterize whether a construction object exists in the road image at the current moment. For example, if a construction object exists, the identification information is 001; if no construction object exists, the identification information is 002.

[0101] The identification information can be determined based on preset judgment rules. For example, if the first detection result indicates that there are road administration facilities in the road image, or the second detection result indicates that there are traffic signs in the road image that meet preset conditions, then the identification information is determined to be 001; if the first detection result indicates that there are no road administration facilities in the road image, and the second detection result indicates that there are no traffic signs in the road image that meet preset conditions, then the identification information is determined to be 002.

[0102] The identification information at the current moment is the final output of the single-frame inference closed loop. Its generation has strict timestamp binding characteristics, corresponding to the end point of the complete processing link of the t-th frame image in the dashcam video stream. Each frame of the road image can correspond to its own identification information.

[0103] S404. Obtain the identification information of a preset number of historical moments before the current moment.

[0104] For example, based on the current time, a preset number of historical times can be determined, and the identification information for each historical time can be obtained. For instance, starting from the current time, N historical times can be determined backwards, where N is the preset number. Each historical time corresponds to its own identification information; therefore, the identification information for each historical time can be obtained.

[0105] S405. Based on the current signage information and a preset number of historical signage information, determine the construction scene information of the road where the vehicle is located.

[0106] For example, by combining the identification information at the current moment with the identification information at a preset number of historical moments, the construction scenario information of the road where the vehicle is located can be determined, that is, whether the road where the vehicle is located is under construction.

[0107] A finite state machine can be pre-configured to make decisions regarding construction scenario information. The finite state machine can be pre-configured with four mutually exclusive states: IDLE, ENTERING, IN_ZONE, and LEAVING. The IDLE state indicates that there is no construction on the road; the ENTERING state indicates that construction is being confirmed; the IN_ZONE state indicates that construction is confirmed; and the LEAVING state indicates that construction is not present. Construction scenario information can be any of these four mutually exclusive states.

[0108] Construction scene information can be determined by the identifier information of the current frame and a preset number (such as the previous 3 frames) of historical identifier information, ensuring that state transitions have temporal continuity constraints and avoiding false broadcasts caused by single-frame jitter. For example, when a vehicle starts, the initial state can be IDLE. If the identifier information at the current moment is 001 and the identifier information at the previous moment is 002, then the IDLE→ENTERING transition is triggered; if the identifier information at the current moment and the previous moment are both 001, then the ENTERING→IN_ZONE transition is triggered; if the identifier information at the current moment is 002 and the identifier information at the previous moment is 001, then the IN_ZONE→LEAVING transition is triggered; if the identifier information at the current moment and the previous moment are both 002, then the LEAVING→IDLE transition is triggered.

[0109] Since momentary rain or fog obscuring the image can lead to misjudgment of identification information in a single frame, this embodiment effectively suppresses noise in single-frame detection by combining identification information from historical moments, thus effectively avoiding false broadcasts. This solves the problem of construction status fluctuations and frequent start-stop broadcasts caused by momentary interference in single-frame detection, ensuring temporal continuity and semantic stability of construction scene information. This provides drivers with reliable, tiered early warning services and offers the cloud platform highly confident structured data on construction events.

[0110] In this embodiment, the construction scene information of the road where the vehicle is located is determined based on the identification information at the current moment and the identification information at a preset number of historical moments. This includes: if the identification information at the current moment is the same as the identification information at a preset number of historical moments, the construction scene information corresponding to the identification information at the current moment is determined based on a preset association relationship; wherein, the preset association relationship represents the association relationship between the identification information and the construction scene information; the method also includes: broadcasting the construction scene information to the user based on a preset broadcasting strategy.

[0111] Specifically, the current moment's identification information is used to characterize whether there is a construction object in the current frame of the road image. The values ​​include, but are not limited to, structured status codes such as "construction object exists", "construction object does not exist", "road maintenance facilities exist but there is no guardrail", and "guardrail exists but there is no sign". The preset number of historical moment identification information refers to the N frames of historical identification information continuously collected and cached before the current moment, where N is a positive integer and can range from 3 to 10.

[0112] The preset association relationships are implemented as a lightweight mapping table in the form of a rule engine, supporting dynamic loading and hot updates, representing the association relationships between identification information and construction scene information. For example, the construction scene information corresponding to identification information 001 is in the IN_ZONE state; the construction scene information corresponding to identification information 002 is in the IDLE state.

[0113] The system determines whether the current identification information is the same as the identification information of a preset number of historical times, i.e., whether N+1 identification information are the same. If so, it determines the construction scenario information corresponding to that identification information based on a preset association relationship. Different construction scenario information corresponds to different preset broadcast strategies. The construction scenario information is broadcast to the user according to the preset broadcast strategy corresponding to the construction scenario information. For example, when the construction scenario information is in the IN_ZONE state, it is broadcast to the user with preset voice content and frequency; when the construction scenario information is in the IDLE state, it is broadcast to the user in the form of a pop-up window.

[0114] If the identification information at the current moment differs from the identification information at a preset number of historical moments, the construction scenario information can be determined based on the number of times the current identification information repeats within the preset number of historical moments. For example, if the number of times the current identification information repeats within the preset number of historical moments is 0, meaning the current identification information is appearing for the first time, the construction scenario information can be determined based on the current identification information. If the number of times the current identification information repeats within the preset number of historical moments is greater than 0, then it is not necessary to determine the construction scenario information. For example, if the current identification information is 001 and appears for the first time, the construction scenario information can be in the ENTERING state; if the current identification information is 002 and appears for the first time, the construction scenario information can be in the LEAVING state.

[0115] In other words, when the vehicle starts, the construction scene information is set to the IDLE state, indicating that no construction has been detected or the construction has been confirmed to be completed. If a construction object is detected during the journey, the vehicle continues to detect the next image frame. If a construction object is detected in multiple consecutive frames, the construction scene information at the current moment is set to the IN_ZONE state. The state between the first frame in which a construction object is detected and the current frame is the ENTERING state. That is, the ENTERING state is the transition from the IDLE state to the IN_ZONE state.

[0116] Starting from the IN_ZONE state, if the disappearance of a construction object is detected for the first time, the state is determined to be LEAVING; if the number of consecutive frames without detected construction reaches N frames, the construction is confirmed to be over, and the state returns to IDLE. That is, the LEAVING state is the transition from the IN_ZONE state to the IDLE state.

[0117] If no construction object is detected during the ENTERING or LEAVING phase, the system can immediately return to the IDLE phase to avoid misjudgment.

[0118] The system broadcasts alerts based on the detected status. For example, upon determining the IN_ZONE status, an alert is immediately triggered and broadcasted periodically. When construction is confirmed to be complete, a completion alert is triggered. Furthermore, the timestamps of the dashcam's video streams can be correlated with the construction status, providing real-time reminders to the driver while simultaneously transmitting construction scene information back to the cloud.

[0119] The benefits of this setup are that, based on time-series consistent identification information, it avoids problems such as false alarms caused by single-frame misdetection, improves the accuracy of construction warnings, and enhances driving safety and the user's driving experience.

[0120] This application provides a method for identifying road construction scenes. It acquires real-time road images of the road where the vehicle is currently located and inputs these images into different preset detection models. Each preset detection model detects a preset category of construction objects. By employing at least two preset detection models in parallel to identify different categories of construction objects, the detection accuracy for different categories of construction objects is improved. The detection results of multiple preset detection models are fused to determine the overall construction scene information, solving problems such as weak generalization ability of single models and susceptibility to interference from single features. This effectively improves the detection accuracy of road construction scenes, thereby enhancing vehicle driving safety and the user's driving experience.

[0121] Figure 5 This is a schematic diagram of the structure of a road construction scene recognition device provided in an embodiment of this application, as shown below. Figure 5 As shown, the road construction scene recognition device 50 provided in this embodiment includes:

[0122] Image acquisition unit 501 is used to acquire a road image of the road where the vehicle is located at the current moment;

[0123] Image detection unit 502 is used to input road images into at least two preset detection models respectively to obtain detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image;

[0124] The scene determination unit 503 is used to determine the construction scene information of the road where the vehicle is located based on the detection results output by each preset detection model; wherein, the construction scene information represents the current construction status of the road.

[0125] In one possible implementation, the preset detection model includes at least a first detection model and a second detection model, wherein the preset category corresponding to the first detection model is road infrastructure, and the preset category corresponding to the second detection model is construction guardrail; the image detection unit 502 includes:

[0126] A first detection module is used to obtain a first detection result based on a road image and a first detection model; wherein the first detection result is the detection result output by the first detection model, and the first detection result characterizes the distribution of road maintenance facilities in the road image; and,

[0127] The second detection module is used to obtain a second detection result based on the road image and the second detection model; wherein, the second detection result is the detection result output by the second detection model, and the second detection result represents the distribution of construction guardrails in the road image.

[0128] In one possible implementation, road infrastructure includes construction barriers; the first detection model includes at least two downsampling convolutional layers; the first detection module is specifically used for:

[0129] The road image is input into the downsampling convolutional layer in the first detection model to obtain the output features corresponding to the downsampling convolutional layer; where the output features are feature vectors representing the road image.

[0130] Based on the output features of each downsampling convolutional layer, construction barriers in the road image are bounded to obtain the number of construction barriers.

[0131] If the number of construction roadblocks exceeds a preset threshold, the first detection result is determined to be the presence of road administration facilities in the road image, and the first detection result is stored in a preset first cache pool.

[0132] In one possible implementation, road infrastructure includes traffic signs; the first detection model includes at least two downsampling convolutional layers; and the first detection module is specifically used for:

[0133] The road image is input into the downsampling convolutional layer in the first detection model to obtain the output features corresponding to the downsampling convolutional layer; where the output features are feature vectors representing the road image.

[0134] If traffic signs are found in the road image based on the output features of each downsampling convolutional layer, then the traffic signs in the road image are classified to determine the type information of the traffic signs; whereby the type information represents the function of the traffic signs.

[0135] If the type information of the traffic sign is determined to meet the preset conditions, the type information of the traffic sign is determined as the first detection result, and the first detection result is stored in the preset first cache pool.

[0136] In one possible implementation, the first detection module is specifically used for:

[0137] The image region of the traffic sign is obtained from the road image, and the image region is scaled based on a preset image range;

[0138] Based on the scaled image area, determine the type information of the traffic sign.

[0139] In one possible implementation, the second detection module is specifically used for:

[0140] The road image is input into the second detection model to determine the area information of the construction guardrail in the road image; where the area information represents the number of pixels occupied by the construction guardrail.

[0141] If the area information of the construction guardrail is greater than the preset area threshold, the second detection result is determined to be that there is a construction guardrail in the road image, and the second detection result is stored in the preset second cache pool.

[0142] In one possible implementation, the scene determination unit 503 includes:

[0143] The identification determination module is used to determine the identification information of the road where the vehicle is located at the current moment based on the detection results output by each preset detection model; wherein, the identification information indicates whether there are construction objects in the road image at the current moment;

[0144] The history determination module is used to obtain the identification information of a preset number of historical moments before the current moment;

[0145] The scene determination module is used to determine the construction scene information of the road where the vehicle is located based on the identification information at the current time and the identification information at a preset number of historical times.

[0146] In one possible implementation, the scene determination module is specifically used for:

[0147] If the identification information at the current moment is the same as the identification information at a preset number of historical moments, then based on the preset association relationship, the construction scene information corresponding to the identification information at the current moment is determined; whereby the preset association relationship represents the association relationship between the identification information and the construction scene information.

[0148] The device also includes a broadcasting unit, which is used to broadcast construction scene information to users based on a preset broadcasting strategy.

[0149] This embodiment provides a road construction scene recognition device that can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0150] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.

[0151] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0152] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0153] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0154] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0155] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0156] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0157] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0158] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0159] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0160] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0161] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0162] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0163] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0164] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0165] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for identifying road construction scenes, characterized in that, include: Obtain a road image of the road where the vehicle is located at the current moment; The road image is input into at least two preset detection models to obtain the detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image; Based on the detection results output by each of the preset detection models, the construction scene information of the road where the vehicle is located is determined; wherein, the construction scene information represents the current construction status of the road.

2. The method according to claim 1, characterized in that, The preset detection model includes at least a first detection model and a second detection model. The preset category corresponding to the first detection model is road infrastructure, and the preset category corresponding to the second detection model is construction guardrail. The road image is input into at least two preset detection models respectively to obtain the detection results output by the preset detection models, including: Based on the road image and the first detection model, a first detection result is obtained; wherein, the first detection result is the detection result output by the first detection model, and the first detection result characterizes the distribution of road maintenance facilities in the road image; and, Based on the road image, a second detection result is obtained using the second detection model; wherein, the second detection result is the detection result output by the second detection model, and the second detection result characterizes the distribution of construction guardrails in the road image.

3. The method according to claim 2, characterized in that, The road infrastructure includes construction barriers; the first detection model includes at least two downsampling convolutional layers; based on the road image and the first detection model, a first detection result is obtained, including: The road image is input into a downsampling convolutional layer in the first detection model to obtain the output features corresponding to the downsampling convolutional layer; wherein, the output features are feature vectors characterizing the road image; Based on the output features corresponding to each downsampling convolutional layer, construction roadblocks in the road image are selected by box selection to obtain the number of construction roadblocks; If the number of construction roadblocks is greater than a preset threshold, the first detection result is determined to be the presence of road administration facilities in the road image, and the first detection result is stored in a preset first cache pool.

4. The method according to claim 2, characterized in that, The road infrastructure includes traffic signs; the first detection model includes at least two downsampling convolutional layers; based on the road image and the first detection model, a first detection result is obtained, including: The road image is input into a downsampling convolutional layer in the first detection model to obtain the output features corresponding to the downsampling convolutional layer; wherein, the output features are feature vectors characterizing the road image; If, based on the output features corresponding to each downsampling convolutional layer, it is determined that the road image contains traffic signs, then the traffic signs in the road image are classified to determine the type information of the traffic signs; wherein, the type information characterizes the function of the traffic signs. If it is determined that the type information of the traffic sign meets the preset conditions, then the type information of the traffic sign is determined as the first detection result, and the first detection result is stored in the preset first cache pool.

5. The method according to claim 4, characterized in that, The traffic signs in the road images are classified to determine their type information, including: The image region of the traffic sign is obtained from the road image, and the image region is scaled based on a preset image range; The type information of the traffic sign is determined based on the scaled image area.

6. The method according to claim 2, characterized in that, Based on the road image and the second detection model, a second detection result is obtained, including: The road image is input into the second detection model to determine the area information of the construction guardrail in the road image; wherein, the area information represents the number of pixels occupied by the construction guardrail; If the area information of the construction guardrail is greater than the preset area threshold, then the second detection result is determined to be that there is a construction guardrail in the road image, and the second detection result is stored in the preset second cache pool.

7. The method according to any one of claims 1-6, characterized in that, Based on the detection results output by each of the preset detection models, the construction scene information of the road where the vehicle is located is determined, including: Based on the detection results output by each of the preset detection models, the identification information of the road where the vehicle is located at the current moment is determined; wherein, the identification information indicates whether there are construction objects in the road image at the current moment; Obtain the identification information for a preset number of historical moments preceding the current moment; Based on the identification information at the current moment and the identification information at the preset number of historical moments, the construction scene information of the road where the vehicle is located is determined.

8. The method according to claim 7, characterized in that, Based on the identification information at the current moment and the identification information at a preset number of historical moments, the construction scene information of the road where the vehicle is located is determined, including: If the identification information at the current moment is the same as the identification information at the preset number of historical moments, then based on the preset association relationship, the construction scene information corresponding to the identification information at the current moment is determined; wherein, the preset association relationship represents the association relationship between the identification information and the construction scene information; The method further includes: broadcasting the construction scene information to the user based on a preset broadcasting strategy.

9. A road construction scene recognition device, characterized in that, include: The image acquisition unit is used to acquire road images of the road where the vehicle is located at the current moment. An image detection unit is used to input the road image into at least two preset detection models to obtain the detection results output by the preset detection models; wherein, the preset detection model is a pre-trained neural network model, the preset detection model is used to identify construction objects of preset categories, and the detection results characterize the distribution of construction objects of preset categories in the road image; The scene determination unit is used to determine the construction scene information of the road where the vehicle is located based on the detection results output by each of the preset detection models; wherein the construction scene information represents the current construction state of the road.

10. An electronic device / computer-readable storage medium / computer program product, characterized in that, The electronic device includes: a memory, a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8; and / or, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-8; and / or, The computer program product includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-8.

Images