A method and system for automatic positioning and digital filing of road assets

Abstract

The application relates to the technical field of digital management of road infrastructure, and discloses a road asset automatic positioning and digital filing method, which comprises the following steps: S1, data preparation, video data of a detection target is acquired by using a detection device to take pictures; S2, video frame picture processing, including target detection and tracking and depth estimation; S3, speed and position estimation, including speed and relative position estimation and absolute position estimation; S4, Kalman filter fusion, multi-source data is input into a linear Kalman filter for processing; and S5, digital filing. The application realizes the spatio-temporal data fusion of multiple sources and different structures by decoupling the target detection, depth estimation and motion analysis modules, and combining the Kalman filter with the inherent geometric prior information of the road environment. The automatic identification, distance measurement, absolute coordinate positioning and digital filing of road assets such as traffic signs can be completed through monocular camera acquisition, and high-precision automatic inspection and digital management of road assets are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital management of road infrastructure, and discloses a method for automatic positioning and digital filing of road assets, in particular to a method for automatic identification, ranging, absolute coordinate positioning and digital filing of road assets such as traffic signs by using vehicle driving cruise video collected by a monocular camera, combining lane marking geometric features and deep learning algorithm. BACKGROUND

[0002] As an important national infrastructure, the safe and efficient operation of road traffic network depends on the effective maintenance and management of various types of road auxiliary facilities (such as traffic signs, guardrails, mileage signs, street lamps, etc.). Establishing a comprehensive, accurate and timely updated digital database of road assets, and clearly defining the category attributes, functional status and precise geographic location of assets in the road network, is the basis for realizing the digital management of the whole life cycle of transportation infrastructure, improving maintenance efficiency and supporting the construction of automatic driving map.

[0003] In the existing application and practice of automatic inspection and digital filing of road assets, manual field investigation is the most traditional method. The manual field investigation method usually drives a patrol vehicle manually, stops or drives at low speed when the target asset is found, uses GPS system, range finder and other devices, and single person or multiple persons cooperate to complete the on-site measurement and recording. This operation mode not only has high labor intensity, low operation efficiency, long time consumption and high operation cost, but also leads to lagging update of asset database, and there is great risk of personal safety hidden danger because the operator needs to be exposed to open traffic flow for a long time, or drives at low speed or even stops on the highway. At the same time, the manual recording process is easily affected by subjective factors, and there may be errors in data entry, with high probability of missed detection, which makes it difficult to ensure the consistency and standardization of road asset information.

[0004] With the development of mobile measurement technology, mobile measurement systems (MMS) based on multi-sensor fusion have become the mainstream high-end solution for acquiring high-precision road data. These systems typically integrate numerous hardware components such as high-beam LiDAR, high-precision Global Navigation Satellite System (GNSS), Inertial Measurement Unit (IMU), and cameras. While they can meet the requirements for high-precision identification, measurement, and positioning during inspections and recording, their high equipment costs and maintenance expenses limit their application in routine, high-frequency daily inspections and their widespread adoption in different regions, especially in areas with low maintenance budgets. Although MMS can acquire 3D point cloud data with centimeter-level accuracy, the post-processing of the massive amounts of point cloud data acquired is extremely complex, involving cumbersome point cloud segmentation, denoising, and feature extraction algorithms, resulting in huge computational resource consumption. More importantly, these systems are highly dependent on GNSS signals. In areas with severe satellite signal obstruction or multipath effects, such as tunnels, deep valleys in mountainous areas, and urban high-rise canyons, positioning signals often drift or are even lost, leading to gaps or large errors in asset positioning, making it difficult to meet the requirements for continuous positioning across the entire road segment.

[0005] In recent years, the rapid advancements in computer vision and deep learning technologies have provided new avenues for low-cost asset detection. Object detection techniques based on convolutional neural networks (such as the YOLO series algorithms) have demonstrated good ability to identify asset categories from vehicle-mounted video images. However, existing monocular vision solutions still face significant technical bottlenecks in inferring three-dimensional spatial location (i.e., "localization") from two-dimensional images. Due to the lack of depth perception capabilities in monocular cameras, there is an inherent scale ambiguity problem, making it difficult to directly obtain accurate distance measurements of target objects. While existing monocular depth estimation networks can predict pixel-level relative depth, their absolute measurement accuracy is often limited by variations in lighting and texture in the training scene, and their generalization ability is weak, resulting in significant errors when directly applied to engineering measurements. Although some research has attempted to utilize binocular stereo vision for distance measurement, this increases hardware costs and calibration complexity. Furthermore, existing pure vision-based localization methods lack effective absolute position correction mechanisms; as the vehicle moves, accumulated localization errors rapidly diverge, reducing the usability of long-distance inspection data.

[0006] Furthermore, developing high-precision asset detection and localization algorithms faces a significant challenge in obtaining the "ground truth." Training deep learning models or validating the accuracy of localization algorithms requires massive amounts of real-world scene data with precise 3D coordinate annotations. Building such datasets in the physical world often necessitates the use of expensive specialized surveying equipment for simultaneous data collection and tedious manual calibration, resulting in extremely high costs and difficulty in covering various complex weather and lighting conditions. Existing research is largely based on limited measured data, lacking systematic validation under controllable, repeatable, and fully known ground truth environments, thus limiting the reliability assessment and iterative optimization of algorithms in practical engineering applications.

[0007] Therefore, there is an urgent need in existing technologies for a low-cost, high-efficiency, and high-precision automated inventory and positioning solution for road assets that can break free from dependence on expensive LiDAR and high-precision GNSS equipment and utilize only widely available ordinary monocular cameras, such as dashcams. How to effectively mine the implicit geometric prior information of the road environment in video images to constrain visual positioning errors and solve the problem of scarce ground truth data in algorithm development are key technical challenges that urgently need to be addressed in this field. Summary of the Invention

[0008] To address the problems of low positioning accuracy, large cumulative error, and reliance on high-cost sensors such as LiDAR in existing monocular vision technologies, this invention provides an automated positioning and digital archiving method for road assets based on monocular video and road geometric features. Its core lies in decoupling target detection, depth estimation, and motion analysis modules, and innovatively utilizing inherent geometric prior information of the road environment—namely, lane marking spacing and mileage signs—as positioning correction anchor points. This is combined with a Kalman filter to achieve multi-source heterogeneous spatiotemporal data fusion. This invention, relying solely on vehicle cruise video collected by dashcams or similar low-cost monocular cameras, automatically identifies, measures, locates, and digitally archives road assets such as traffic signs, achieving high-precision automated inspection and digital management of road assets.

[0009] The technical solution provided by this invention is as follows: On the one hand, this invention provides a method for automated location and digital archiving of road assets, including the following steps: S1. Data preparation: Use detection equipment to capture and acquire video data of the detection targets, including road assets and road signs. S2. Video frame processing, including the following steps: S21. Target detection and tracking: Record the target type and detection time based on the video data obtained in S1. S22. Depth estimation: Based on the target detection and tracking results obtained in S21, a pre-trained depth estimation network is used to convert image pixels into depth measurements. The depth estimation network is based on a visual Transformer architecture and outputs a pixel-level depth map through an encoder-decoder structure. For the depth prediction results, the depth value of the center point of the target bounding box is taken as a representative value to obtain the distance between the detection device and the detected target. S3, Velocity and Position Estimation: This step corrects positioning errors using prior environmental geometric information and includes two sub-steps: S31. Speed ​​and relative position estimation: The video data acquired in S1 is processed by grayscale and binarization to extract lane markings. A virtual detection line, i.e. a virtual road marking, is set up. The lane marking spacing specified by the mandatory national standard is used as a periodic reference to calculate the relative displacement. The speed and relative position information of the current image are analyzed through the known geometric features of the road markings. S32. Absolute position estimation: Extract position information from road signs and combine it with the extracted relative position information to convert it into the absolute position of the detection equipment. S4. Kalman filter fusion: The multi-source data, consisting of noisy raw data, the physical geometric relationship between the detection equipment and the road signs, and the speed and absolute position information of the detection equipment, is input into a linear Kalman filter. The data is integrated and noise is reduced to obtain unstructured video data of the location of the road signs. S5. Digital archiving: Transform the unstructured video data output from the S4 process into a structured road asset database.

[0010] Preferably, S21 target detection and tracking is used to identify road asset categories and output bounding box coordinates to achieve continuous tracking of targets in video sequences. S21 adopts a YOLO11s convolutional neural network model, with input data being a monocular cruise video frame sequence. The output results will be used as input for subsequent depth estimation and directly provided to subsequent linear Kalman filtering as one of the observation data sources.

[0011] Preferably, S21 uses a deep learning model to independently process visual semantic information. This model extracts image features through a convolutional neural network and uses non-maximum suppression to eliminate duplicate detection boxes. For the detection of distant and small targets, the YOLO11s convolutional neural network model integrates a feature pyramid network in the neck layer, upsamples the deep semantic feature map and fused it with the shallow high-resolution feature map to enhance the feature representation capability of small targets. When the detected target is occluded, its position is predicted based on the historical motion trajectory to maintain tracking continuity.

[0012] Preferably, S22 depth estimation is used to solve the scale blur problem of monocular vision and outputs the absolute distance from the target to the detection device. To suppress depth prediction noise, S22 adopts a two-stage filtering mechanism: spatial filtering to eliminate image noise and temporal filtering to suppress depth fluctuations caused by vehicle vibration carrying the detection device. The input of this step is the cropped image within the target detection bounding box; the output is a scalar depth value, representing the absolute distance from the target to the monocular camera. The output result will be used as the input of the subsequent Kalman filter to calculate the position of the target in the road network coordinate system.

[0013] Preferably, in step S31, speed and relative position estimation utilizes the geometric features of lane markings to calculate the motion parameters of the vehicle carrying the detection equipment; lane markings are extracted through image processing, and the lane marking spacing specified by mandatory national standards is used as a periodic reference to calculate the relative displacement; combined with the time information in the video frame sequence, the current driving speed of the vehicle carrying the detection equipment is calculated; the input of this step is the original video frame; the output is the speed and relative driving distance of the vehicle carrying the detection equipment.

[0014] Preferably, S32 absolute position estimation is used to identify road signs and provide absolute coordinate anchor points. By cropping the images of road signs in the target detection results, only the bounding boxes of road signs in the target detection are retained. The content of the road signs is parsed using OCR optical character recognition technology, and rule verification is combined to ensure the reliability of recognition. The recognition results are mapped to the road network coordinate system as a global position reference. The road signs serve as high-precision absolute coordinate anchor points, providing continuous positioning correction. The input for this step is the original video frame. When there is a detection and recognition result, the output is the absolute coordinate anchor point displayed by the road sign; otherwise, it is a null value.

[0015] Preferably, S4 Kalman filter fusion is used to dynamically fuse multi-source observation data and output the absolute position of the asset; by constructing a state space model, a state vector containing the asset position and velocity is defined; the state transition equation is based on the assumption of uniform motion and describes the kinematic characteristics of the vehicle carrying the detection equipment; the multi-source data composed of the observation data obtained from S21 target detection and tracking, the motion analysis data obtained from S32 motion and position estimation, and the predicted depth data obtained from S22 depth estimation are used as observation inputs, and the prediction and update steps are executed iteratively through the Kalman filter; the Kalman gain dynamically balances the prediction and observation weights and outputs the optimized position estimate.

[0016] Preferably, in S5, converting the unstructured video data output from the S4 process into a structured road asset database specifically includes: S51. Data structuring: For each road asset object that is successfully detected, tracked and located, the system generates a structured record containing the following fields; Asset ID: A unique identifier generated based on the tracking sequence; Asset Category: The category determined by the target detection module (e.g., "Speed ​​Limit 120", "No Parking", etc.); Absolute station number: The absolute mileage location of the road network calculated by the Kalman filter fusion module (e.g., K1024+530m); Real-world images: Save several images corresponding to the target detection results as a schematic diagram of this road asset; Timestamp: The specific time when the asset was captured; S52, Anomaly Handling and Manual Review Interface: The system automatically marks low-confidence segments as "Pending Verification"; the filing software interface provides a manual review function, where operators can click on a record to jump to the corresponding original video segment for quick confirmation or correction; S53. Database storage: Structured data is stored in a database to form searchable and updatable electronic archives of road assets, supporting the management and analysis of the road maintenance management system.

[0017] On the other hand, the present invention provides an automated road asset location and digital filing system for implementing the above-mentioned automated road asset location and digital filing method. The system includes: Video Input Module: Receives video data of the target captured by the detection equipment, and inputs the video data (i.e., the monocular video stream of the detection equipment's patrol) after timestamp alignment processing into the video frame processing module. The video frame processing module includes three parallel processing sub-modules: target detection and tracking, depth estimation, and motion and position estimation. The video input module distributes data to these three sub-modules of the video frame processing module. The target detection and tracking submodule records the target type and detection time in the video data acquired by the video input module, processes the video frames using a deep learning model, and outputs the asset category and bounding box coordinates; it maintains continuous tracking of the target in the video sequence through a trajectory matching algorithm. Depth estimation submodule: Uses the DepthAnythingV2 pre-trained model for depth estimation, outputs the absolute distance from the detected target to the detection device, performs depth prediction on the cropped region based on the bounding box output by the target detection, and outputs the absolute distance from the target to the monocular camera after noise suppression processing; Motion and position estimation submodule: includes lane marking analysis unit and road sign recognition unit; Lane marking analysis unit: Through image preprocessing, marking extraction and displacement calculation, outputs the motion status of the vehicle carrying the detection equipment; Road sign recognition unit: It analyzes the content of road signs through optical character recognition technology and outputs absolute coordinate anchor points.

[0018] Velocity and position estimation module: includes velocity and relative position estimation submodule and absolute position estimation submodule; The speed and relative position estimation submodule analyzes the current screen's speed and relative position information using the known geometric features of road markings; The absolute position estimation submodule extracts position information from road signs and combines it with the extracted relative position information to convert it into the absolute position of the detection device. Kalman filter fusion module: Receives multi-source observation data, performs state prediction and update, and outputs the absolute position coordinates of the asset; Digital filing module: Stores the location results in a structured manner in the asset database, forming a searchable electronic file.

[0019] Compared with the prior art, the beneficial effects of the present invention are: Significant cost-effectiveness: It only uses widely available monocular cameras, eliminating the need for expensive equipment such as LiDAR or high-precision positioning systems, greatly reducing hardware costs and data processing difficulty, and making it suitable for high-frequency daily inspections in different regions.

[0020] Positioning accuracy meets engineering requirements: By dynamically fusing environmental geometric prior constraints and Kalman filtering, the scale ambiguity problem of monocular vision is effectively solved, achieving positioning accuracy that meets engineering standards.

[0021] Improved efficiency and safety: Fully automated processing significantly improves inspection efficiency; no need to slow down or stop for measurement, eliminating personnel safety risks.

[0022] High environmental adaptability: The modular and decoupled design enables the system to adapt to complex road conditions such as straight roads and curves; it does not rely on a positioning system, ensuring stable operation in areas with signal obstruction such as tunnels.

[0023] High technical reliability: Environmental geometric prior information provides a continuous correction mechanism, effectively suppressing cumulative positioning errors and ensuring the availability of long-distance inspection data.

[0024] Modular decoupling design: The three functional modules of target detection, depth estimation, and motion analysis are processed independently to avoid error propagation caused by information coupling and improve system robustness; Environmental geometric prior constraints: Using the standard lane marking spacing (6-meter line + 9-meter gap) specified by the mandatory national standard as a relative displacement counter, and the mileage sign as an absolute coordinate anchor point, the inherent scale ambiguity problem of monocular vision is effectively corrected. Dynamic data fusion mechanism: Construct a state-space model and dynamically integrate multi-source observation data through a Kalman filter to suppress noise and correct accumulated positioning errors. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below.

[0026] Figure 1 A simplified flowchart of the method for automated location and digital filing of road assets provided by this invention; Figure 2 A flowchart illustrating the steps involved in video frame processing. Figure 3This is a flowchart illustrating the steps involved in velocity and position estimation. Figure 4 This is a schematic diagram of the process of fusing road assets using Kalman filtering to obtain the final location; Figure 5 This is a schematic diagram illustrating the geometric relationship between the detection equipment and the road assets.

[0027] Figure 6 A simplified diagram of the automated location and digital filing system for road assets provided by this invention; Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0029] In one embodiment, the present invention provides a method for automated location and digital archiving of road assets, such as... Figure 1 As shown, it includes the following steps: S1. Data preparation, including the geometric relationship between the detection equipment and road assets, such as... Figure 5 As shown, the detection device in this application is a monocular camera, which can be vehicle-mounted. The detection device is used to capture and acquire video data of the detection target. The detection target includes road assets and road signs. Here, the road signs mainly include mileage markers with mileage data. S2, video frame processing, such as Figure 2 As shown, it includes the following steps: S21. Target detection and tracking: Record the target type and detection time based on the video data obtained in S1. S21 target detection and tracking is used to identify road asset categories and output bounding box coordinates to achieve continuous tracking of targets in video sequences. S21 adopts the YOLO11s convolutional neural network model. The input data is a monocular cruise video frame sequence. The output results will be used as the input for subsequent depth estimation and directly provided to the subsequent linear Kalman filter as one of the observation data sources.

[0030] The specific implementation steps for S21 are as follows: 1. Object Detection The input to this module is a sequence of cruise video frames captured by a monocular camera. The video frames are first fed into the YOLO11s object detection model, which needs to be trained on the road asset dataset of the corresponding category.

[0031] The YOLO11s model's network structure includes the following key components: The backbone network employs the CSPDarknet53 architecture, consisting of several input convolutional layers, a spatial pyramid pooling (SPP) module, and a cross-stage part (CSP) module. Through multi-level feature extraction, the backbone network outputs feature maps at different scales: 80×80, 40×40, and 20×20, corresponding to 1 / 8, 1 / 16, and 1 / 32 resolution of the input image, respectively.

[0032] The Neck network integrates a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN) to upsample deep semantic feature maps and then concatenate and fuse them with shallow high-resolution feature maps. The FPN part contains 3 upsampling layers and 3 feature fusion layers, while the PAN part contains 3 downsampling layers and 3 feature fusion layers, forming a bidirectional feature pyramid structure to enhance multi-scale feature representation capabilities.

[0033] The head network consists of three detectors, each making predictions at different scales. Each detector contains five convolutional layers to predict bounding box coordinates, class probabilities, and confidence scores. The output format is [x1, y1, x2, y2, confidence, class1, class2, ...], where (x1, y1) and (x2, y2) represent the coordinates of the top-left and bottom-right corners of the bounding box, respectively, confidence represents the confidence in the presence of the target, and class1, class2, etc., represent the probabilities of each class.

[0034] During the detection process, a confidence threshold of 0.80 and an intersection-over-union (IoU) threshold of 0.8 were set to filter out low-quality detection results. The model output format is as follows: Bounding box coordinates: [x1, y1, x2, y2], representing the coordinates of the top left and bottom right corners of the detected target in the image; Category Labels: Integer identifiers that map to the detected asset categories based on the structure built from the dataset; Confidence level: A floating-point value in the range [0,1], representing the reliability of the detection result.

[0035] 2. Target tracking For targets in video sequences, a dual-constraint tracking algorithm based on IoU and motion velocity is adopted. The specific steps are as follows: First, sort the target detection results by timestamp. For each detected target in the current frame, calculate the intersection-over-union (IoU) with all detected targets in the previous frame, as shown in formula (1).

[0036] (1); In equation (1), A and B are the regions of the two bounding boxes on the image, respectively; This represents the area of ​​intersection between the two regions; This represents the area of ​​the union of the two regions.

[0037] set up The minimum threshold is 0.5, when If the value is greater than this threshold, the two bounding boxes are considered to belong to the same target. If multiple bounding boxes exist... If all values ​​are greater than the threshold, then select... Match the value with the largest value.

[0038] When the target is occluded or temporarily disappears, an assisted tracking method is used to match the target based on motion velocity constraints. The moving velocity of the target center point in the current frame and the previous frame is calculated as shown in formula (2).

[0039] (2); In equation (2), Movement speed; The Euclidean distance between the center points of the target in the two frames; This is the time interval between two frames.

[0040] The maximum speed threshold is set to 20 pixels / frame. When the calculated speed is less than this threshold, a pseudo IoU value is used for matching, as shown in formula (3).

[0041] (3); In equation (3), This is a pseudo-IoU value; To adjust the parameter, the value is set to 0.1; The target moving speed is calculated using formula (2).

[0042] A minimum pseudo-IoU threshold of 0.3 is set. When the pseudo-IoU value is greater than this threshold, the two detection boxes are considered to belong to the same target. During tracking, the minimum number of consecutive frames for the target to appear is set to 3 frames to filter out noise detection.

[0043] 3. Category Determination For each tracking sequence, the confidence score of the detection results in each frame is accumulated, as shown in formula (4).

[0044] (4); In equation (4), Here, n is the cumulative confidence score for category c, and n is the length of the tracking sequence. Let be the confidence level of the i-th frame. This is an indicator function that takes the value 1 when the category of the i-th frame is c, and 0 otherwise.

[0045] The category with the highest cumulative confidence level is used as the final category for the tracking sequence to ensure the stability of the category determination. When a mileage marker is detected, the system will specifically mark the target for use by subsequent modules.

[0046] S22. Depth Estimation: Based on the target detection and tracking results obtained in S21, depth estimation is performed using the DepthAnythingV2 pre-trained model. This step is used to obtain the distance between the detection device and the detected target. Based on the bounding box output by the target detection, depth prediction is performed on the cropped region. After noise suppression processing, the absolute distance from the target to the monocular camera is output. This addresses the scale blur problem of monocular vision and outputs the absolute distance from the target to the detection device. The pre-trained depth estimation network converts image pixels into depth measurements. The depth estimation network is based on a visual Transformer architecture and outputs pixel-level depth maps through an encoder-decoder structure. For the depth prediction results, the depth value of the center point of the target bounding box is taken as the representative value because it has high consistency in road assets with planar interfaces, such as road signs.

[0047] To suppress depth prediction noise, S22 employs a two-stage filtering mechanism: spatial filtering eliminates image noise, and temporal filtering suppresses depth fluctuations caused by vehicle vibrations carrying the detection equipment. The input to this step is a cropped image within the target detection bounding box; the output is a scalar depth value, representing the absolute distance from the target to the monocular camera. The output result will be used as the input to the subsequent Kalman filter to calculate the target's position in the road network coordinate system.

[0048] The specific implementation steps of S22 are as follows: 1. Depth estimation model configuration The depth estimation model employs DepthAnythingV2, using the DINOv2 visual Transformer as the encoder and a dense prediction Transformer as the decoder. This pre-trained model is built upon the ViT-Small encoder with a feature dimension of 64 and is trained using the VKITTI dataset. The model's input is a sequence of cruise video frames, and its output is the absolute depth information of the images in the cruise video frame sequence, i.e., the metric distance from each pixel in the image to the monocular camera, in meters.

[0049] The structure of the DepthAnythingV2 model is as follows: Encoder: A visual Transformer pre-trained based on DINOv2, containing 12 Transformer layers, each with 12 attention heads, for a total of 21.5M parameters. The encoder transforms the input image into a multi-scale feature representation and outputs feature maps at four different scales.

[0050] Decoder: Dense Prediction Transformer, consisting of 4 Transformer layers, each with 8 attention heads. The decoder receives multi-scale features from the encoder output and generates a depth map with the same resolution as the input image through upsampling and feature fusion operations.

[0051] The regression head contains three convolutional layers that convert the features output by the decoder into depth values. The first layer uses a 1×1 convolution to reduce the number of channels to 64, the second layer uses a 3×3 convolution to refine the features, and the third layer uses a 1×1 convolution to output a single-channel depth map.

[0052] 2. Deep Prediction The input to this module is a sequence of cruise video frames in which the detected object appears during target detection. For each detected target, the depth value of the center point of the bounding box is taken as the representative distance from the target to the monocular camera, as shown in formula (5).

[0053] (5); In equation (5), D represents the depth from the target to the monocular camera; D is the depth map. The coordinates are the center point of the target bounding box.

[0054] 3. Depth Correction To address the systematic errors in the depth prediction results, a correction method based on prior knowledge is employed. Based on road environment characteristics, the default settings are: road sign height 3.0 meters, monocular camera installation height 1.5 meters, and sign lateral offset distance 6.125 meters. These parameters can be adjusted according to the actual scenario.

[0055] Based on geometric relationships, calculate the scaling factor between the true depth and the predicted depth. As shown in formula (6).

[0056] (6); In equation (6), The height of the sign; The height of the monocular camera; The vertical angle between the monocular camera and the sign.

[0057] The corrected depth value is shown in formula (7).

[0058] (7); To suppress depth prediction noise, a two-stage filtering mechanism is employed: Spatial filtering: A 3×3 Gaussian filter is applied to the depth map, with the standard deviation of the filter kernel set to 1.0, to eliminate image noise. Temporal filtering: Apply a moving average filter to the depth values ​​of 5 consecutive frames to obtain the filtered depth values. As shown in formula (8).

[0059] (8); In equation (8), N is the size of the filtering window, which is 5. For the first The depth value of the frame.

[0060] S3. Speed ​​and Position Estimation: This step utilizes prior environmental geometric information from road markings to analyze the current image's speed and relative position. It obtains the absolute position by analyzing information from mileage markers and provides absolute coordinate anchor points to correct positioning errors. This includes the following steps: S31, velocity and relative position estimation, such as Figure 3 As shown, the main focus is on lane marking processing and statistics. The video data acquired by S1 is processed by grayscale and binarization to extract lane markings. A virtual detection line, i.e., a virtual road marking, is set up. The lane marking spacing (6-meter line + 9-meter gap) specified by the mandatory national standard is used as a periodic reference to calculate relative displacement. The geometric features of the lane markings are used to calculate the motion parameters of the vehicle carrying the detection equipment. Combined with the time information in the video frame sequence, the current speed of the vehicle carrying the detection equipment is calculated. The input for this step is the original video frame; the output is the speed and relative distance traveled by the vehicle carrying the detection equipment.

[0061] The specific implementation steps of S31 are as follows: 1. Lane marking analysis: This function analyzes lane markings by preprocessing the image, extracting the markings, and calculating displacement. It outputs the motion state of the vehicle carrying the detection equipment. The specific implementation steps are as follows: 1.1 Image Preprocessing The input RGB format original video frames are converted to grayscale as shown in formula (9).

[0062] (9); In equation (9), R, G, and B are the red, green, and blue channel values ​​of the image, respectively. The values ​​of the above parameters are all in the range of [0, 255], representing the processed grayscale values.

[0063] The grayscale image is binarized, and the grayscale threshold is set to 180, as shown in formula (10).

[0064] (10); 1.2 Lane Marking Extraction The specific steps for extracting lane marking contours using a contour detection algorithm are as follows: (1) Perform morphological closing operation on the binarized image to eliminate small holes and breaks; (2) Use the Canny edge detection algorithm to extract edges; (3) Apply contour detection algorithm to obtain all closed contours.

[0065] The extracted contours are filtered by parameters such as width, height, and area, retaining only those that meet the constraints set by formulas (11), (12), and (13): (11); (12); (13); In equations (11)-(13), w, h, and A represent the width, height, and area of ​​the extracted contour, respectively. The corresponding index is the lower limit of the threshold. This represents the upper limit of the threshold.

[0066] 1.3 Displacement Calculation A virtual detection line is set in the image. When the bottom and top center points of the effective lane marking contour extracted by S3112 pass through the detection line, a timestamp is recorded and counted. Based on the timestamp and count results, the relative position of the vehicle carrying the detection equipment is calculated, as shown in formula (14).

[0067] (14); In equation (14), Let be the relative position at time t; These are the number of virtual detection lines passing through the top and bottom center points of the effective lane marking outline, respectively. For the first The driving distance corresponding to each lane marking.

[0068] According to the mandatory national standard GB5768.3, the lane markings of expressways adopt the design of "6-meter line + 9-meter gap", hence formula (15).

[0069] (15); Formula (15) indicates that when the top center point of the effective lane marking profile passes through the virtual detection line, the relative displacement is 6 meters; when the bottom center point of the effective lane marking profile passes through the virtual detection line, the relative displacement is 9 meters.

[0070] 1.4 Speed ​​Calculation The original velocity is calculated based on the change in relative displacement and the corresponding timestamp, as shown in formula (16).

[0071] (16); In equation (16), Let be the initial velocity at time t; For time intervals.

[0072] To suppress fluctuations caused by data noise, a moving average window is used for smoothing, as shown in formula (17).

[0073] (17); In equation (17), denoted as t, where is the smooth speed at time t; N is the window size.

[0074] Using linear interpolation, the velocity at any given time is calculated as shown in formula (18).

[0075] (18); In equation (18), and For the closest time t, and there exists and The time point of the data.

[0076] S32. Absolute position estimation: Extract position information from road signs (image cropping, OCR processing), and combine it with the extracted relative position information to convert it into the absolute position of the detection device.

[0077] S32 is used to identify road signs and provide absolute coordinate anchor points. By cropping the image of objects classified as "road signs" from the target detection results, only the bounding boxes of those categories are retained. OCR (Optical Character Recognition) technology is used to parse the road sign content, and rule-based verification ensures recognition reliability. The recognition results are mapped to the road network coordinate system as a global position reference. Road signs serve as high-precision absolute coordinate anchor points, providing continuous positioning correction. The input for this step is the original video frame; when there are detection and recognition results, the output is the absolute coordinate anchor point displayed by the road sign; otherwise, it is a null value.

[0078] The specific implementation steps of S32 are as follows: 1. Road sign recognition: In this embodiment, the recognition of mileage signs is mainly performed. The content of the road signs is analyzed using optical character recognition technology, and the absolute coordinate anchor points are output. The specific implementation steps are as follows: 1.1 Target Screening From the results of the S21 target detection, target objects classified as mileage signs are selected, and the images within the bounding boxes of each target are cropped.

[0079] 1.2 OCR Recognition and Processing For the cropped image in step 1.1 above, use PaddleOCR for optical character recognition, set to English character mode to reduce interference from other characters, and set the confidence threshold to 0.85.

[0080] The recognition results are post-processed to retain only the numeric characters, i.e., "0123456789", and remove other characters.

[0081] The recognition result of this step is a number character such as "1001" or "1002", which represents the mileage markers on the highway at the 1001km, 1002km, etc., and their locations. Other values ​​are deduced in the same way.

[0082] Road signs may be detected multiple times in consecutive frames. Trajectory matching is performed using timestamps and location information, and the recognition result with the highest confidence is selected as the final output.

[0083] 1.3 Absolute Position Calculation The recognition result of the mileage sign is used as the anchor point of the absolute position, as in formula (19).

[0084] (19); In equation (19), To determine the absolute position at a specific time; To determine the absolute position of the anchor point at that moment, This refers to the mileage sign value displayed in the OCR recognition results.

[0085] The relative displacement and absolute position anchor point obtained from the above steps are combined, and the absolute position is calculated using formula (20).

[0086] (20); In equation (20), The absolute mileage value identified from the most recently passed mileage marker; This represents the relative displacement accumulated from lane marking counts at the current moment. This represents the relative displacement at the moment of passing the mileage marker anchor point.

[0087] S4, Kalman filter fusion, such as Figure 4 As shown, in this embodiment, this step is used for the localization of road signs. Specifically, the multi-source data, consisting of noisy raw data, the physical geometric relationship between the detection device and the road sign, and the speed and absolute position (cropping, OCR processing) information of the detection device, is input into a linear Kalman filter. After Kalman filtering, the data is fused and integrated and noise is reduced to obtain unstructured video data of the location of the road sign.

[0088] S4 is used to dynamically fuse multi-source observation data and output the absolute position of the asset. A state-space model is constructed, defining a state vector containing the asset's position and velocity. The state transition equation, based on the assumption of uniform motion, describes the kinematic characteristics of the vehicle carrying the detection equipment. Multi-source data, consisting of observation data obtained from target detection and tracking in S21, motion analysis data obtained from motion and position estimation in S22, and predicted depth data obtained from depth estimation in S23, is used as the observation input. A Kalman filter iteratively executes prediction and update steps. The Kalman gain dynamically balances the prediction and observation weights, outputting an optimized position estimate. When road sign anchors are valid, the system performs absolute coordinate hard correction, effectively suppressing accumulated positioning errors. This step receives the outputs of all the above steps, fuses them, and outputs structured position data, completing the closed loop of the entire processing flow. The output is the asset's category and absolute position coordinates.

[0089] Kalman filter fusion is the core computation for achieving high-precision positioning. However, depth information acquired by monocular vision sensors is affected by illumination, image resolution, and model accuracy, resulting in significant random observation noise. Direct use of this information can lead to abrupt changes in positioning results. Furthermore, while the speed of vehicles carrying detection equipment can be calculated from lane markings, measurement errors still exist. This invention employs a linear Kalman filter, utilizing the continuity constraint of vehicle motion with detection equipment, to perform spatiotemporal fusion of visual observation data and lane marking motion data, achieving optimized estimation of the relative position and relative speed of road assets.

[0090] S4 specifically includes the following processes: S41. Geometric Space Transformation and Observation Preprocessing The depth estimation module outputs metric depth, which is the straight-line Euclidean distance from the optical center of the monocular camera to the center of the target object (such as a road sign). However, road asset localization requires the target's longitudinal position in the road network coordinate system. Therefore, the straight-line depth is first converted into a longitudinal horizontal distance using geometric relationships.

[0091] Based on the appendix Figure 2 Based on the geometric relationships shown, construct the transformation formula (21).

[0092] (twenty one); In equation (21), For a moment The visual observation longitudinal distance, in meters; The output of the depth estimation model at time [time] The predicted depth value of the target center point; The prior value for the center height of the road sign to be detected; The installation height of the vehicle-mounted monocular camera is the prior value. The prior value is the lateral distance between the center of the target object and the centerline of the vehicle carrying the detection equipment.

[0093] In addition, the vehicle speed carrying the detection equipment is obtained via S41 at any given time. .

[0094] S42. Establish the system state-space model To describe the dynamic relationship between road assets and the monocular camera, the system's state vector at time t is defined as shown in Equation (22).

[0095] (twenty two); In equation (22), For a moment The estimated longitudinal distance of the target asset relative to the monocular camera; For a moment The actual approach speed of the vehicle carrying the detection equipment relative to the target asset.

[0096] Considering that the video frame rate of devices such as dashcams is generally no less than 24 frames and the inter-frame interval is no more than about 0.042 seconds, it can be assumed that the vehicle carrying the detection equipment moves at a constant speed in a straight line during this extremely short time interval. Based on this assumption, the discrete-time state transition equations are constructed as shown in formulas (23) and (24).

[0097] (twenty three); (twenty four); In equations (23) and (24), This is the state transition matrix, used to describe the evolution of the system state over time; Indicates time and The time interval between moments; Let be the process noise vector, with a mean of 0 and a covariance of . The Gaussian distribution is used to compensate for non-uniform motion errors caused by slight acceleration / deceleration and road bumps during the actual driving of the vehicle carrying the detection equipment; the first row of the state transition matrix... This means the current distance = the distance at the previous moment - (velocity × time interval). The negative sign here indicates that the vehicle carrying the detection equipment is moving forward, and the distance between the target and the monocular camera decreases; the second row of the state transition matrix... This indicates that the current velocity is equal to the velocity at the previous moment, which is the assumption of the uniform velocity model.

[0098] S43. Establishment of the observation equation The mapping relationship between sensor measurements and the actual state of the system is described using observation equations, such as Equation (25).

[0099] (25); In equation (25), For a moment The observation vector contains measurement data from visual and lane marking analysis, as shown in Equation (26).

[0100] (26); H is the observation matrix. Since it only involves two state variables, distance and velocity, H is the identity matrix, as shown in formula (27).

[0101] (27); To measure the noise vector, it follows a mean of 0 and a covariance of . The Gaussian distribution is shown in Equation (28), which describes the uncertainty of the sensor itself.

[0102] (28); In equation (28), The variance of the depth estimate, This is the variance of the speed calculated using lane markings.

[0103] S44, Recursive Filtering Iteration For the same target being tracked continuously in a video sequence, the Kalman filter performs a cyclical operation of "prediction" and "update" in each frame.

[0104] Step 1: Prediction Phase. This phase, based on the optimal estimate from the previous moment and the physical motion model, calculates the current state and uncertainties. The posterior estimate predicts the current time. The prior state.

[0105] The predicted state estimate is represented by formula (29).

[0106] (29); In equation (29), Indicates at time Prior state estimate, index This indicates that the value is based on time. and previous information regarding time The predictions made; Indicates the previous moment The posterior state estimate is the result of the previous recursive loop.

[0107] The prediction error covariance is characterized by formula (30), which transfers the uncertainty of the previous time step to the current time step and adds process noise. In order to control the range of uncertainty.

[0108] (30); In equation (30), Indicates time The prior error covariance matrix, which describes the predicted state. The degree of deviation from the actual state, with the diagonal elements of the matrix representing the variance of the longitudinal distance prediction and the variance of the velocity prediction, respectively; Indicates the previous moment The posterior error covariance matrix, i.e. the uncertainty remaining after the previous update; The transpose of the state transition matrix ; This is the process noise covariance matrix, which reflects the uncertainty of acceleration.

[0109] Step 2: Correction Phase. This phase incorporates the actual observation data at the current moment. The prior results obtained in the prediction stage are corrected.

[0110] Calculate Kalman gain , as in formula (31). Kalman gain characterizes a tradeoff factor that determines whether the filter places more trust in "model predictions" or "sensor observations." This is relevant when observation noise... When smaller, Increasing the filter makes it more confident in the observed values; conversely, decreasing the filter makes it more confident in the predicted values.

[0111] (31); Update the state estimate as shown in Equation (32). This is the core output of the filter. The residual represents the difference between the actual observed value and the predicted value. The filter uses the residual to correct the prior estimate and obtain an optimized posterior estimate.

[0112] (32); Update the error covariance, as shown in formula (33).

[0113] (33); In equation (33), I is the identity matrix. This step updates the uncertainty of the system; due to the introduction of new observation information, the uncertainty... The uncertainty is usually less than that of the forecast. .

[0114] S45. Absolute Position Calculation and Output When target tracking ends (i.e., the target leaves the video frame or is obscured and lost) or at any time When the position needs to be output, extract the state vector. The first component That is, the relative longitudinal distance after filtering and smoothing.

[0115] Ultimate road asset absolute position The calculation is as shown in formula (34).

[0116] (34); In equation (34), For a moment The absolute position of the monocular camera is calculated from the relative displacement between the mileage marker anchor point and the lane markings.

[0117] Through the above Kalman filtering fusion process, random noise in monocular depth estimation can be effectively eliminated, and road asset positioning with centimeter-level accuracy can be achieved by utilizing environmental geometric features.

[0118] S5. Digital Archiving: This process transforms the unstructured video data output from the S4 workflow into a structured road asset database. Specifically, it includes: S51. Data structuring: For each road asset object that is successfully detected, tracked and located, the system generates a structured record containing the following fields; Asset ID: A unique identifier generated based on the tracking sequence; Asset Category: The category determined by the target detection module (e.g., "Speed ​​Limit 120", "No Parking", etc.); Absolute station number: The absolute mileage location of the road network calculated by the Kalman filter fusion module (e.g., K1024+530m); Real-world images: Save several images corresponding to the target detection results as a schematic diagram of this road asset; Timestamp: The specific time when the asset was captured.

[0119] S52, Anomaly Handling and Manual Review Interface: The system automatically marks low confidence levels (e.g., classification confidence level below 0.6 or location variance). Records that are too large are marked as "Pending Verification". The filing software interface provides a manual review function, where operators can click on a record to jump to the corresponding original video segment for quick confirmation or correction.

[0120] S53. Database storage: Structured data is stored in relational databases such as MySQL and geographic information system (GIS) databases to form searchable and updatable electronic archives of road assets, supporting the management and analysis of road maintenance management systems.

[0121] In another embodiment, the present invention provides an automated road asset location and digital archiving system for implementing the above-mentioned automated road asset location and digital archiving method. The system includes: Video Input Module: Receives video data of the target captured by the detection equipment, and inputs the video data (i.e., the monocular video stream of the detection equipment's patrol) after timestamp alignment processing into the video frame processing module. The video frame processing module includes two parallel processing sub-modules: a target detection and tracking sub-module and a depth estimation sub-module. The video input module distributes data to the two sub-modules of the video frame processing module. The target detection and tracking submodule records the target type and detection time in the video data acquired by the video input module, processes the video frames using a deep learning model, and outputs the asset category and bounding box coordinates; it maintains continuous tracking of the target in the video sequence through a trajectory matching algorithm. Depth estimation submodule: Uses the DepthAnythingV2 pre-trained model for depth estimation, outputs the absolute distance from the detected target to the detection device, performs depth prediction on the cropped region based on the bounding box output by the target detection, and outputs the absolute distance from the target to the monocular camera after noise suppression processing; The speed and position estimation module parses the current screen speed and relative position information, obtains the absolute position by parsing the information shown on the mileage marker, and provides absolute coordinate anchor points. The speed and position estimation module includes two sub-modules: a speed and relative position estimation sub-module and an absolute position estimation sub-module. Speed ​​and relative position estimation submodule: Lane marking analysis unit: Through image preprocessing, lane markings are extracted, and the known geometric features of the road markings are used to analyze the speed and relative position information of the current image, and output the motion status of the vehicle carrying the detection device; The absolute position estimation submodule extracts position information from road signs and combines it with the extracted relative position information to convert it into the absolute position of the detection device. Kalman filter fusion module: Receives multi-source observation data, performs state prediction and update, and outputs the absolute position coordinates of the asset.

[0122] Digital filing module: Stores the location results in a structured manner in the asset database, forming a searchable electronic file.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

Claims

1. A method for automated location and digital archiving of road assets, characterized in that, Includes the following steps: S1. Data preparation: Use detection equipment to capture and acquire video data of the detection targets, including road assets and road signs. S2. Video frame processing, including the following steps: S21. Target detection and tracking: Record the target type and detection time based on the video data obtained in S1. S22. Depth estimation: Based on the target detection and tracking results obtained in S21, a pre-trained depth estimation network is used to convert image pixels into depth measurements. The depth estimation network is based on a visual Transformer architecture and outputs a pixel-level depth map through an encoder-decoder structure. For the depth prediction results, the depth value of the center point of the target bounding box is taken as a representative value to obtain the distance between the detection device and the detected target. S3, Velocity and Position Estimation: This step corrects positioning errors using prior environmental geometric information and includes two sub-steps: S31. Speed ​​and relative position estimation: The video data acquired in S1 is processed by grayscale and binarization to extract lane markings. A virtual detection line, i.e. a virtual road marking, is set up. The lane marking spacing specified by the mandatory national standard is used as a periodic reference to calculate the relative displacement. The speed and relative position information of the current image are analyzed through the known geometric features of the road markings. S32. Absolute position estimation: Extract position information from road signs and combine it with the extracted relative position information to convert it into the absolute position of the detection equipment. S4. Kalman filter fusion: The multi-source data, consisting of noisy raw data, the physical geometric relationship between the detection equipment and the road signs, and the speed and absolute position information of the detection equipment, is input into a linear Kalman filter. The data is integrated and noise is reduced to obtain unstructured video data of the location of the road signs. S5. Digital archiving: Transform the unstructured video data output from the S4 process into a structured road asset database.

2. The method for automated location and digital archiving of road assets according to claim 1, characterized in that, S21 target detection and tracking is used to identify road asset categories and output bounding box coordinates to achieve continuous tracking of targets in video sequences. S21 adopts the YOLO11s convolutional neural network model. The input data is a monocular cruise video frame sequence. The output results will be used as the input for subsequent depth estimation and directly provided to the subsequent linear Kalman filter as one of the observation data sources.

3. The method for automated location and digital archiving of road assets according to claim 2, characterized in that, S21 employs a deep learning model to independently process visual semantic information. This model extracts image features through a convolutional neural network and uses non-maximum suppression to eliminate duplicate detection boxes. For the detection of distant, small targets, the YOLO11s convolutional neural network model integrates a feature pyramid network in the neck layer, upsampling the deep semantic feature map and concatenating it with the shallow high-resolution feature map to enhance the feature representation capability of small targets. When the detected target is occluded, its position is predicted based on the historical motion trajectory to maintain tracking continuity.

4. The method for automated location and digital archiving of road assets according to claim 1, characterized in that, S22 depth estimation addresses the scale blur problem in monocular vision, outputting the absolute distance from the target to the detection device. To suppress depth prediction noise, S22 employs a two-stage filtering mechanism: spatial filtering eliminates image noise, and temporal filtering suppresses depth fluctuations caused by vehicle vibrations carrying the detection device. The input to this step is a cropped image within the target detection bounding box; the output is a scalar depth value representing the absolute distance from the target to the monocular camera. This output will serve as the input to the subsequent Kalman filter for calculating the target's position in the road network coordinate system.

5. The method for automated location and digital archiving of road assets according to claim 1, characterized in that, S31 speed and relative position estimation utilizes lane marking geometric features to calculate the motion parameters of the vehicle carrying the detection equipment; lane markings are extracted through image processing, and the lane marking spacing specified by mandatory national standards is used as a periodic reference to calculate the relative displacement; combined with the time information in the video frame sequence, the current driving speed of the vehicle carrying the detection equipment is calculated. The input for this step is the original video frame; the output is the speed and relative distance traveled by the vehicle carrying the detection equipment.

6. The method for automated location and digital archiving of road assets according to claim 1, characterized in that, S32 absolute position estimation is used to identify road signs and provide absolute coordinate anchor points. By cropping the image of the target detection result that is classified as a road sign, only the bounding box of the target detection category is retained. The content of the road sign is parsed using OCR optical character recognition technology, and rule verification is combined to ensure the reliability of recognition. The recognition results are mapped to the road network coordinate system as a global position reference; road signs serve as high-precision absolute coordinate anchor points, providing continuous positioning correction; the input for this step is the original video frame; when there is a detection and recognition result, the output is the absolute coordinate anchor point shown by the road sign, otherwise it is a null value.

7. The method for automated location and digital archiving of road assets according to claim 1, characterized in that, The S4 Kalman filter fusion is used to dynamically fuse multi-source observation data and output the absolute position of the asset. A state-space model is constructed to define a state vector containing the asset's position and velocity. The state transition equation is based on the assumption of uniform motion and describes the kinematic characteristics of the vehicle carrying the detection equipment. The multi-source data, consisting of observation data obtained from target detection and tracking in S21, motion analysis data obtained from motion and position estimation in S22, and predicted depth data obtained from depth estimation in S23, is used as the observation input. The prediction and update steps are executed iteratively through the Kalman filter. The Kalman gain dynamically balances the prediction and observation weights to output an optimized position estimate.

8. The method for automated location and digital archiving of road assets according to claim 1, characterized in that, In S5, the transformation of unstructured video data output from the S4 process into a structured road asset database specifically includes: S51. Data structuring: For each road asset object that is successfully detected, tracked and located, the system generates a structured record containing the following fields; Asset ID: A unique identifier generated based on the tracking sequence; Asset Category: The category determined by the target detection module; Absolute station number: The absolute mileage position of the road network calculated by the Kalman filter fusion module; Real-world images: Save several images corresponding to the target detection results as a schematic diagram of this road asset; Timestamp: The specific time when the asset was captured; S52, Anomaly Handling and Manual Review Interface: The system automatically marks low-confidence segments as "Pending Verification"; the filing software interface provides a manual review function, where operators can click on a record to jump to the corresponding original video segment for quick confirmation or correction; S53. Database storage: Structured data is stored in a database to form searchable and updatable electronic archives of road assets, supporting the management and analysis of the road maintenance management system.

9. A road asset automated positioning and digital filing system, used to implement the method described in any one of claims 1-8, characterized in that, The system includes: Video Input Module: Receives video data of the target captured by the detection equipment, and inputs the video data (i.e., the monocular video stream of the detection equipment's patrol) after timestamp alignment processing into the video frame processing module. The video frame processing module includes two parallel processing sub-modules: a target detection and tracking sub-module and a depth estimation sub-module. The video input module distributes data to the two sub-modules of the video frame processing module. The target detection and tracking submodule records the target type and detection time in the video data acquired by the video input module, processes the video frames using a deep learning model, and outputs the asset category and bounding box coordinates; it maintains continuous tracking of the target in the video sequence through a trajectory matching algorithm. Depth estimation submodule: Uses the DepthAnythingV2 pre-trained model for depth estimation, outputs the absolute distance from the detected target to the detection device, performs depth prediction on the cropped region based on the bounding box output by the target detection, and outputs the absolute distance from the target to the monocular camera after noise suppression processing; Speed ​​and relative position estimation submodule: Through image preprocessing, lane markings are extracted, and the known geometric features of the road markings are used to analyze the speed and relative position information of the current image, and output the motion status of the vehicle carrying the detection device. The absolute position estimation submodule extracts position information from road signs and combines it with the extracted relative position information to convert it into the absolute position of the detection device. Kalman filter fusion module: Receives multi-source observation data, performs state prediction and update, and outputs the absolute position coordinates of the asset; Digital filing module: Stores the location results in a structured manner in the asset database, forming a searchable electronic file.

Images