Method and device for fusing depth ranging of road lane line and post number plate, equipment and medium

Abstract

The application discloses a kind of fusing road lane line and pole number plate depth ranging method, device, equipment and medium, it is related to data processing technical field, including: based on binocular camera acquisition road image and pre-processing, pre-processing image is input into multi-task fusion perception network containing shared encoder, time convolution unit etc., extract lane line pixel coordinate set and pole number plate information;It is obtained by multi-constraint joint ranging double depth value, then adaptive weighted fusion obtains target depth value, finally time fusion inference obtains smooth depth estimation value and outputs semantic position information containing stake number, distance. Distance is measured by fusing lane line structure and pole number plate semantic information, without additional sensor, low in cost, easy to deploy, improve the accuracy and robustness of road stake depth ranging in complex environment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, equipment and medium for depth ranging that integrates road lane lines and station markers. Background Technology

[0002] Currently, road marker depth ranging primarily employs a combination of visual feature recognition and distance measurement techniques. One approach involves using visual recognition technology to detect targets and parse text on the road markers, then combining this with camera parameters or simple stereo matching algorithms to calculate the depth distance between the markers and the camera. Another approach extracts the geometric features of the road lane lines separately, using the structural constraints of the lane lines to extrapolate spatial distances and assist in road location estimation. Furthermore, some ranging schemes rely on GPS signals for auxiliary positioning or employ expensive sensors such as LiDAR for depth ranging. Overall, none of these technical approaches effectively integrate lane line structural features with the semantic information of the road markers.

[0003] However, relying solely on mileage marker recognition for depth ranging is susceptible to environmental factors such as shooting angle deviations, image blurring, lighting changes, and partial occlusion of mileage markers. Furthermore, the lack of spatial geometric constraints leads to low ranging accuracy and poor result stability. Using only lane line features for distance estimation cannot directly correlate with the absolute mileage information of the mileage markers, making it difficult to achieve globally accurate positioning of road mileage markers; only relative distance calculations are possible. Simultaneously, existing technologies do not combine prior structural information such as lane line parallelism and spacing with the semantic information of the mileage markers to form a dual constraint, resulting in insufficient robustness of depth ranging and significant ranging errors in complex road environments such as fog, rain, and snow. Moreover, solutions relying on GPS signals are prone to signal loss in scenarios such as tunnels and tall buildings, while solutions using additional sensors such as LiDAR significantly increase deployment costs, making large-scale application in existing traffic monitoring equipment and vehicle systems difficult.

[0004] Therefore, how to integrate the geometric constraints of road lane lines with the semantic information of road station markers to improve the accuracy and robustness of road station depth measurement in complex environments has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, device, equipment, and medium for depth measurement that integrates road lane lines and station markers, aiming to solve the technical problem of how to improve the accuracy and robustness of road station depth measurement in complex environments.

[0006] To achieve the above objectives, this application proposes a depth ranging method that integrates road lane lines and station markers, comprising:

[0007] Acquire road image sequences using a binocular camera;

[0008] The road image sequence is preprocessed to obtain a preprocessed image;

[0009] The preprocessed image is input into a preset multi-task fusion perception network for processing to obtain the pixel coordinate set of lane lines and station number information. The preset multi-task fusion perception network includes a shared encoder, a temporal convolutional unit, an environmental condition modulation module, and a multi-task decoder. The shared encoder adopts a multi-stage residual structure and embeds a temporal convolutional unit at the output end. The multi-task decoder adopts a dual-branch parallel structure, which includes a lane line branch and a station number branch. The lane line branch adopts an upsampling-convolution-upsampling U-shaped structure, and the station number branch adopts a feature pyramid fusion structure. The environmental condition modulation module consists of a fully connected network and a batch normalization layer.

[0010] Based on the set of pixel coordinates of the lane lines and the information of the station markers, a first depth value and a second depth value are calculated using a multi-constraint joint depth ranging algorithm.

[0011] The target depth value is obtained by adaptive weighted fusion of the first depth value and the second depth value.

[0012] The target depth value is subjected to time-consistent spatiotemporal fusion and correlation reasoning to obtain a smooth depth estimate and output semantic location information, wherein the semantic location information includes station information, depth distance, timestamp and environmental state.

[0013] In one embodiment, the step of preprocessing the road image sequence to obtain a preprocessed image includes:

[0014] Collect environmental perception data, perform timestamp alignment and standardization processing on the environmental perception data, and generate an environmental coding vector;

[0015] Based on the environment encoding vector, an adaptive adjustment instruction for exposure parameters is generated;

[0016] Based on the adaptive adjustment command for the exposure parameters, the exposure time and gain of the binocular camera are dynamically adjusted to obtain a road image sequence with adjusted parameters.

[0017] Based on the weather score in the environmental coding vector, environmentally-aware image enhancement processing is performed on the road image sequence after parameter adjustment to obtain a preprocessed image.

[0018] In one embodiment, the step of performing environment-aware image enhancement processing on the parameter-adjusted road image sequence based on the weather score in the environment coding vector to obtain a preprocessed image includes:

[0019] Obtain a weather score, which includes a fog concentration score, a rain / snow score, and a visibility score;

[0020] When the fog concentration score of the weather rating is greater than the preset fog concentration threshold, a lightweight defogging algorithm is enabled on the road image sequence after parameter adjustment to obtain the first preprocessed image;

[0021] When the rain and snow score of the weather score is greater than the preset rain and snow threshold, a preset rain pattern suppression filter is applied to the road image sequence after parameter adjustment to obtain a second preprocessed image.

[0022] When the visibility score of the weather rating is lower than the preset visibility threshold, the ROI region of the road image sequence after parameter adjustment is reduced to obtain a third preprocessed image;

[0023] When the fog concentration score of the weather score is greater than the preset fog concentration threshold, and the rain and snow score of the weather score is greater than the preset rain and snow threshold, the lightweight defogging algorithm and the preset rain pattern suppression filter are sequentially applied to the road image sequence after parameter adjustment to obtain the fourth preprocessed image.

[0024] When the fog concentration score of the weather rating is greater than the preset fog concentration threshold and the visibility score of the weather rating is lower than the preset visibility threshold, a lightweight defogging algorithm and a reduction of the ROI region are sequentially applied to the road image sequence after parameter adjustment to obtain the fifth preprocessed image.

[0025] When the rain / snow score of the weather rating is greater than the preset rain / snow threshold and the visibility score of the weather rating is lower than the preset visibility threshold, the preset rain pattern suppression filter and the ROI region are applied sequentially to the road image sequence after parameter adjustment to obtain the sixth preprocessed image.

[0026] When the fog concentration score of the weather score is greater than the preset fog concentration threshold, the rain and snow score of the weather score is greater than the preset rain and snow threshold, and the visibility score of the weather score is lower than the preset visibility threshold, the lightweight defogging algorithm, the preset rain pattern suppression filter, and the ROI region are sequentially applied to the road image sequence after parameter adjustment to obtain the seventh preprocessed image.

[0027] Select the corresponding image from the first preprocessed image, the second preprocessed image, the third preprocessed image, the fourth preprocessed image, the fifth preprocessed image, the sixth preprocessed image, and the seventh preprocessed image as the preprocessed image.

[0028] In one embodiment, the step of processing the preprocessed image into a preset multi-task fusion perception network to obtain the pixel coordinate set of lane lines and marker information includes:

[0029] The preprocessed image of the current processing frame and the preprocessed images of the previous preset number of frames are combined to form a temporal image sequence, which is then input into the shared encoder of the preset multi-task fusion perception network to obtain multi-scale spatial features.

[0030] Temporal fusion of the multi-scale spatial features is performed using the temporal convolutional unit to obtain temporally enhanced features;

[0031] The temporal enhancement features are input into the environmental condition modulation module, and the modulation parameters of lane line branches and station mark branches are generated according to the environmental coding vector to obtain the modulated features.

[0032] The modulated features are input into the lane line branch of the multi-task decoder, and lane line instance segmentation and curve fitting are performed on the modulated features to obtain the set of pixel coordinates of the lane lines.

[0033] The post number branch of the multi-task decoder is used to perform post number target detection and text recognition on the modulated features to obtain post number information containing bounding box coordinates and post number text.

[0034] In one embodiment, the step of calculating the first depth value and the second depth value based on the pixel coordinate set of the lane line and the station marker information using a multi-constraint joint depth ranging algorithm includes:

[0035] Input the set of pixel coordinates of the lane lines into a pre-calibrated inverse perspective transformation matrix to complete the conversion from pixel coordinates to bird's-eye view coordinates, and obtain the set of bird's-eye view coordinates of the lane lines.

[0036] A three-dimensional geometric model of the lane lines is constructed by performing polynomial fitting on the bird's-eye view coordinate set of the lane lines.

[0037] Based on the three-dimensional geometric model of the lane line and the bounding box coordinates in the station sign information, the world coordinate system coordinates of the bottom midpoint of the station sign are determined to obtain the initial first depth value.

[0038] The bounding box coordinates in the station number information are extracted as the region of interest. The binocular image pairs acquired by the binocular camera are subjected to regional stereo matching to obtain the disparity map of the station number region.

[0039] The effective disparity pixels of the disparity map of the mileage marker area are filtered and the mean is calculated to obtain the effective disparity value of the mileage marker area.

[0040] The effective parallax value of the station marker area is combined with the camera intrinsic parameters and the binocular baseline distance to calculate the initial second depth value;

[0041] The initial first depth value and the initial second depth value are respectively checked for reasonableness, and invalid values ​​that exceed the preset distance threshold are removed to obtain the first depth value and the second depth value.

[0042] In one embodiment, the step of adaptively weighting and fusing the first depth value and the second depth value to obtain the target depth value includes:

[0043] Extract the lane line detection confidence score corresponding to the first depth value and the station marker detection confidence score corresponding to the second depth value to obtain two sets of detection confidence scores.

[0044] The historical ranging error variances of the first depth value and the second depth value within a preset time window are statistically analyzed to obtain two sets of error variance data.

[0045] The two sets of detection confidence values ​​and the two sets of error variance data are input into the weight calculation model, and the first weight of the first depth value and the second weight of the second depth value are obtained through normalization processing.

[0046] Multiply the first depth value by the first weight to obtain the first weighted depth value, and multiply the second depth value by the second weight to obtain the second weighted depth value.

[0047] The first weighted depth value and the second weighted depth value are summed to obtain the initial fusion depth value;

[0048] The initial fusion depth value is subjected to neighborhood smoothing filtering to obtain the target depth value.

[0049] In one embodiment, the step of performing temporally consistent spatiotemporal fusion and correlation reasoning on the target depth value to obtain a smooth depth estimate and output semantic location information includes:

[0050] Multi-target tracking is performed on the set of lane line pixel coordinates, station number information and corresponding target depth values ​​of multiple consecutive frames, and a unique identifier is assigned across frames to obtain an identifiable spatiotemporal perception dataset.

[0051] Calculate the lateral distance between the location of the mileage marker corresponding to the mileage marker information and the pixel coordinate set of each lane line, establish a spatiotemporal association pair between the mileage marker and its corresponding lane line, and obtain the associated depth dataset;

[0052] Extract the perception quality score corresponding to the environmental perception data, and adjust the process noise covariance of the adaptive Kalman filter according to the perception quality score to obtain the parameter-calibrated Kalman filter.

[0053] The target depth value in the associated depth dataset is input into the parameter-calibrated Kalman filter for time-series smoothing filtering to obtain a smoothed depth estimate.

[0054] By integrating the smoothed depth estimate, the station number text in the station number information, the timestamp of image acquisition, and the environmental perception data, a structured location data set is generated;

[0055] The structured location data set is formatted and labeled with confidence to obtain semantic location information containing station information, depth distance, timestamp and environmental status.

[0056] Furthermore, to achieve the above objectives, this application also proposes a depth ranging device that integrates road lane lines and mileage markers, the depth ranging device comprising:

[0057] The acquisition module is used to acquire road image sequences using a binocular camera;

[0058] The preprocessing module is used to preprocess the road image sequence to obtain a preprocessed image;

[0059] The fusion perception module is used to process the preprocessed image input into a preset multi-task fusion perception network to obtain the pixel coordinate set of lane lines and station number information. The preset multi-task fusion perception network includes a shared encoder, a temporal convolutional unit, an environmental condition modulation module, and a multi-task decoder. The shared encoder adopts a multi-stage residual structure and embeds a temporal convolutional unit at the output end. The multi-task decoder adopts a dual-branch parallel structure, which includes a lane line branch and a station number branch. The lane line branch adopts an upsampling-convolution-upsampling U-shaped structure, and the station number branch adopts a feature pyramid fusion structure. The environmental condition modulation module consists of a fully connected network and a batch normalization layer.

[0060] The joint ranging module is used to calculate, based on the set of pixel coordinates of the lane lines and the station marker information, a multi-constraint joint depth ranging algorithm to obtain a first depth value and a second depth value.

[0061] The weighted calculation module is used to perform adaptive weighted fusion based on the first depth value and the second depth value to obtain the target depth value;

[0062] The results module is used to perform time-consistent spatiotemporal fusion and correlation reasoning on the target depth value to obtain a smooth depth estimate and output semantic location information, wherein the semantic location information includes station information, depth distance, timestamp and environmental state.

[0063] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the depth ranging method for fusing road lane lines and station markers as described above.

[0064] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the depth ranging method for fusing road lane lines and station markers as described above.

[0065] This application utilizes binocular cameras to acquire and preprocess road images. The preprocessed images are then input into a multi-task fusion perception network containing a shared encoder and temporal convolutional units to extract lane line pixel coordinates and station marker information. Dual depth values ​​are obtained through multi-constraint joint ranging, followed by adaptive weighted fusion to obtain the target depth value. Finally, temporal fusion inference yields a smooth depth estimate, outputting semantic location information including station markers and distance. By fusing lane line structure and station marker semantic information for ranging, no additional sensors are required, resulting in low cost, easy deployment, and improved accuracy and robustness of road station depth ranging in complex environments. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart illustrating the first embodiment of the depth ranging method integrating road lane lines and station markers in this application;

[0068] Figure 2 This is a flowchart illustrating the second embodiment of the depth ranging method integrating road lane lines and station markers in this application;

[0069] Figure 3 This is a schematic diagram of the module structure of the depth measuring device that integrates road lane lines and station markers, according to the first embodiment of the depth measuring method integrating road lane lines and station markers of this application.

[0070] Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the depth ranging method that integrates road lane lines and station markers in the embodiments of this application.

[0071] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0072] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0073] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0074] Currently, road marker depth ranging primarily employs a combination of visual feature recognition and distance measurement techniques. One approach involves using visual recognition technology to detect targets and parse text on the road markers, then combining this with camera parameters or simple stereo matching algorithms to calculate the depth distance between the markers and the camera. Another approach extracts the geometric features of the road lane lines separately, using the structural constraints of the lane lines to extrapolate spatial distances and assist in road location estimation. Furthermore, some ranging schemes rely on GPS signals for auxiliary positioning or employ expensive sensors such as LiDAR for depth ranging. Overall, none of these technical approaches effectively integrate lane line structural features with the semantic information of the road markers.

[0075] However, relying solely on mileage marker recognition for depth ranging is susceptible to environmental factors such as shooting angle deviations, image blurring, lighting changes, and partial occlusion of mileage markers. Furthermore, the lack of spatial geometric constraints leads to low ranging accuracy and poor result stability. Using only lane line features for distance estimation cannot directly correlate with the absolute mileage information of the mileage markers, making it difficult to achieve globally accurate positioning of road mileage markers; only relative distance calculations are possible. Simultaneously, existing technologies do not combine prior structural information such as lane line parallelism and spacing with the semantic information of the mileage markers to form a dual constraint, resulting in insufficient robustness of depth ranging and significant ranging errors in complex road environments such as fog, rain, and snow. Moreover, solutions relying on GPS signals are prone to signal loss in scenarios such as tunnels and tall buildings, while solutions using additional sensors such as LiDAR significantly increase deployment costs, making large-scale application in existing traffic monitoring equipment and vehicle systems difficult.

[0076] Therefore, how to integrate the geometric constraints of road lane lines with the semantic information of road station markers to improve the accuracy and robustness of road station depth measurement in complex environments has become an urgent problem to be solved.

[0077] Based on the above, this application provides a depth ranging method that integrates road lane lines and marker posts, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the depth ranging method integrating road lane lines and station markers in this application.

[0078] In this embodiment, the depth ranging method that integrates road lane lines and marker posts includes steps S10 to S60:

[0079] Step S10: Acquire road image sequences using a binocular camera.

[0080] Specifically, a binocular camera module is fixedly installed at the center of the inner side of the vehicle's windshield. The baseline distance between the left and right cameras is set to 12-20cm and the optical axes are kept parallel. The left and right frames are strictly synchronized and exposed through a hardware trigger signal. The image sequence in front of the road is continuously acquired at a resolution of 1920×1080 and a frame rate of 25fps. Each frame of the image is marked with a microsecond-level timestamp and written into the synchronization buffer.

[0081] Step S20: Preprocess the road image sequence to obtain a preprocessed image.

[0082] It should be noted that step S20 includes: First, collecting environmental perception data, performing timestamp alignment and standardization on the environmental perception data, and generating an environmental coding vector. It is important to understand that environmental perception data refers to various types of data related to the road environment obtained from different sensing channels, covering data such as light intensity, weather type, fog concentration, rain and snow intensity, and visibility. This data comprehensively reflects the real-time environmental state of the road scene, providing data support for subsequent adaptive environmental processing. The environmental coding vector is a vector form of data formed by transforming the standardized environmental perception data through a specific coding method, which can condense multi-dimensional environmental information into a single vector expression. Specifically, three independent acquisition threads are launched in parallel to acquire multi-source data in real time: the first thread reads physical measurements from an external illuminance meter, rain sensor, and haze concentration sensor at a frequency of 100Hz via an RS485 interface; the second thread performs high-speed parallel calculations on the current frame of the binocular camera to extract intrinsic image statistics such as average brightness, global contrast, and Laplacian gradient magnitude; and the third thread listens for CAN bus broadcast messages through the vehicle's OBD interface to capture vehicle context signals such as wiper gear position, headlight on / off status, and vehicle speed. All data is stamped with a microsecond-level hardware timestamp and written to a circular buffer at the moment of acquisition. Next, multi-data sets with the same timestamp are extracted from the buffer and analyzed for physical... Sensor readings are processed using Kalman filtering to eliminate transient jitter. Image statistics are averaged using a 5-frame sliding window to suppress single-frame anomalies. Vehicle status signals are encoded as discrete level identifiers, which are then mapped to a continuous 0-1 interval based on the physical meaning of each indicator and preset empirical thresholds, forming a dimensionless standardized feature vector. Finally, the standardized feature vector is input into a pre-trained lightweight SVM classifier. Through the nonlinear mapping of the radial basis function kernel function, the association patterns of multimodal information are comprehensively inferred, and a structured environment encoding vector is output. This vector includes three quantitative dimensions: light intensity level, rain, fog, and snow weather score, and image quality index, providing a real-time environmental status representation for subsequent adaptive exposure adjustment and image enhancement strategy selection.

[0083] Next, based on the environment encoding vector, an adaptive adjustment instruction for exposure parameters is generated. Then, based on the adaptive adjustment instruction, the exposure time and gain of the stereo camera are dynamically adjusted to obtain a road image sequence with adjusted parameters.

[0084] Finally, based on the weather score in the environment coding vector, environmentally-aware image enhancement processing is performed on the parameter-adjusted road image sequence to obtain a preprocessed image. Further, a weather score is obtained, which includes a fog concentration score, a rain / snow score, and a visibility score. When the fog concentration score is greater than a preset fog concentration threshold, a lightweight dehazing algorithm is applied to the parameter-adjusted road image sequence to obtain a first preprocessed image. When the rain / snow score is greater than a preset rain / snow threshold, a preset rain ripple suppression filter is applied to the parameter-adjusted road image sequence to obtain a second preprocessed image. When the visibility score is lower than a preset visibility threshold, the ROI region of the parameter-adjusted road image sequence is reduced to obtain a third preprocessed image. When both the fog concentration score and the rain / snow score are greater than a preset fog concentration threshold, a lightweight dehazing algorithm and a preset rain ripple suppression filter are sequentially applied to the parameter-adjusted road image sequence to obtain a fourth preprocessed image. When both the fog concentration score and the visibility score are greater than a preset fog concentration threshold, a lightweight dehazing algorithm and a preset rain ripple suppression filter are sequentially applied to the parameter-adjusted road image sequence to obtain a fourth preprocessed image. When a visibility threshold is set, a lightweight dehazing algorithm and ROI region reduction are sequentially applied to the parameter-adjusted road image sequence to obtain a fifth preprocessed image. When the rain / snow score of the weather rating is greater than a preset rain / snow threshold and the visibility score of the weather rating is lower than a preset visibility threshold, a preset rain streak suppression filter and ROI region reduction are sequentially applied to the parameter-adjusted road image sequence to obtain a sixth preprocessed image. When the fog concentration score of the weather rating is greater than a preset fog concentration threshold and the rain / snow score of the weather rating is greater than a preset rain / snow threshold and the visibility score of the weather rating is lower than a preset visibility threshold, a lightweight dehazing algorithm, a preset rain streak suppression filter, and ROI region reduction are sequentially applied to the parameter-adjusted road image sequence to obtain a seventh preprocessed image. The corresponding image is selected as the preprocessed image from the first, second, third, fourth, fifth, sixth, and seventh preprocessed images.Specifically, the weather score component in the environmental encoding vector E is analyzed, and three sub-dimensions—fog concentration score, rain / snow score, and visibility score—are extracted. When the fog concentration score is greater than 0.7, a lightweight dehazing algorithm based on dark channel prior is applied to the parameter-adjusted road image sequence. Atmospheric light values ​​and transmittance maps are estimated through guided filtering, and atmospheric scattering model inversion is performed on each frame to recover the clear scene radiation, resulting in the first pre-processed image after dehazing. When the rain / snow score is greater than 0.6, a frequency domain rain streak suppression filter is applied, and the image is transformed to the Fourier domain. Directional band-stop filtering is used to eliminate high-frequency components of rain streaks, and the image is restored through inverse transformation, resulting in the second pre-processed image after rain removal. When the visibility score is less than 0.4, the ROI region is automatically reduced to the center of the original image. The image is focused on the road surface within 50 meters in front of the vehicle, and the signal-to-noise ratio is improved by reducing the processing resolution to obtain the third pre-processed image after focusing. When multiple conditions are met simultaneously, fog, rain, and ROI are processed in sequence. That is, for fog and rain coexisting scenes, fog is removed first and then rain is removed; for low visibility foggy days, fog is removed first and then ROI is reduced; for low visibility rainy days, rain is removed first and then ROI is reduced; for extreme weather, fog removal, rain removal, and ROI focusing are performed in sequence. Finally, based on the combination of real-time weather scores, the corresponding result is selected from seven processing paths as the pre-processed image output to ensure that the key road features (lane line edges, marker outlines) of the image sequence under different severe weather conditions remain identifiable and stable for subsequent processing.

[0085] Step S30: The preprocessed image is input into a preset multi-task fusion perception network for processing to obtain the pixel coordinate set of the lane lines and the station marker information.

[0086] It should be noted that the preset multi-task fusion perception network includes a shared encoder, a temporal convolutional unit, an environmental condition modulation module, and a multi-task decoder. The shared encoder adopts a multi-stage residual structure and embeds a temporal convolutional unit at the output. The multi-task decoder adopts a dual-branch parallel structure, which includes a lane line branch and a station sign branch. The lane line branch adopts an upsampling-convolution-upsampling U-shaped structure, and the station sign branch adopts a feature pyramid fusion structure. The environmental condition modulation module consists of a fully connected network and a batch normalization layer.

[0087] Specifically, firstly, the preprocessed image of the current frame and its four preceding historical frames are stacked chronologically to form a five-frame temporal image sequence, which is then input into the multi-stage residual structure of the shared encoder. Each stage consists of multiple convolutional residual blocks, and each residual block is a bottleneck structure (1×1, 3×3, 1×1 convolution). Temporal convolutional units (ConvLSTM) are embedded at the outputs of the four stages (conv2_x, conv3_x, conv4_x, conv5_x) of ResNet-50. The spatial feature map of each stage is used as the input at the current time step, and temporal fusion is performed by combining it with the hidden state at the previous time step to output a multi-scale feature pyramid that enhances temporal consistency. Secondly, the environment encoding vector E is input into the fully connected network of the environment condition modulation module. Two 256-dimensional fully connected layers are used to generate scaling factors and biases for lane line branches and station marker branches, respectively, for the shared encoder. The temporal enhancement features output by the generator are normalized by a batch normalization layer and then subjected to a channel-by-channel affine transformation. This enhances edge gradient features in the lane line branch during foggy weather and suppresses high-frequency noise features in the station sign branch during rainy weather. The modulated features are then input in parallel into a dual-branch decoder. The lane line branch uses a U-shaped structure, recovering resolution through a two-stage skip connection of transposed convolution upsampling - 3×3 convolution - transposed convolution upsampling. A 1×1 convolution segmentation head outputs a pixel-level lane line probability map, and clustering extracts the pixel coordinate set for each lane line. The station sign branch uses a feature pyramid fusion structure, aligning multi-scale features through 1×1 convolution and then horizontally connecting them from top to bottom. A YOLO detection head outputs the station sign bounding box coordinates. Finally, the detected bounding box regions are cropped and normalized before being input into an improved lightweight PaddleOCR. In the text detection stage, an enhanced lightweight DB text detection network is used as the starting point for improvement. Its core improvement lies in the introduction of a multi-scale feature enhancement neck network. Specifically, after extracting multi-level features in the backbone network MobileNetV3-Small, consecutive 1×1 convolutional layers are used to perform channel dimensionality reduction and alignment on feature maps from different levels. Then, a spatial attention mechanism is used to weight the contributions of feature maps at different scales. Finally, element-wise addition is used to fuse the features with a single 3×3 depthwise separable convolution. This improvement enables the detection network to more sensitively capture station number text regions at different scales, from clear large characters in the foreground to blurred small characters in the background, significantly improving the accuracy and recall of bounding box regression under complex lighting and partial occlusion. In the text correction and feature extraction stages, firstly, after the spatial transformation network, a mini-network consisting of two residual blocks is used, embedding asymmetric convolution decompositions (such as decomposing a 3×3 convolution into a concatenation of 1×3 and 3×1 convolutions). This structure can more effectively capture the horizontal and vertical stroke features of characters without significantly increasing the number of parameters, enhancing the model's ability to discriminate subtle differences in character structure. Secondly, the standard convolutional layers in CRNN are replaced with Ghost modules.The Ghost module generates eigenmaps using a small number of regular convolutions, and then generates more feature maps through linear transformation 1×1 convolutions, thereby reducing the computational complexity of the convolutional layers while maintaining the ability to express similar features. In the sequence modeling and recognition stage, a lightweight gated recurrent unit (GRU) and self-attention hybrid module are used. Specifically, a GRU with fewer layers and more streamlined hidden units is used instead of LSTM to capture long-term dependencies in the sequence, reducing computational cost. A lightweight single-head self-attention layer is introduced on the output sequence of the GRU. This layer maps the GRU output sequence to Query, Key, and Value through 1×1 convolutions, calculates attention weights, and allows features at each position in the sequence to directly interact with features at all positions within the sequence, thus more accurately modeling the contextual relationships between characters and outputting station number text information. The final network output is formed by integrating the lane line pixel coordinates, station number bounding box coordinates, and station number text.

[0088] Step S40: Calculate the first depth value and the second depth value using a multi-constraint joint depth ranging algorithm based on the pixel coordinate set of the lane lines and the station marker information.

[0089] It should be noted that step S40 includes: inputting the set of pixel coordinates of the lane lines into a pre-calibrated inverse perspective transformation matrix to complete the conversion from pixel coordinates to bird's-eye view coordinates, obtaining a set of bird's-eye view coordinates for the lane lines; performing polynomial fitting on the set of bird's-eye view coordinates for the lane lines to construct a three-dimensional geometric model of the lane lines; determining the world coordinate system coordinates of the bottom midpoint of the station sign based on the three-dimensional geometric model of the lane lines and the bounding box coordinates in the station sign information, obtaining an initial first depth value; extracting the bounding box coordinates in the station sign information as the region of interest, performing regional stereo matching on the binocular image pairs acquired by the binocular camera, obtaining a disparity map of the station sign region; filtering effective disparity pixels and calculating the mean value on the disparity map of the station sign region, obtaining an effective disparity value for the station sign region; calculating the effective disparity value of the station sign region in combination with camera intrinsic parameters and binocular baseline distance, obtaining an initial second depth value; performing a rationality check on the initial first depth value and the initial second depth value respectively, removing invalid values ​​that exceed a preset distance threshold, obtaining the first depth value and the second depth value.

[0090] It's important to understand that the set of pixel coordinates for lane lines refers to the coordinate data of each point on the lane line extracted from the image, located on the image plane. These coordinates are represented in pixels and can completely describe the position and shape information of the lane lines in the image. The pre-calibrated inverse perspective transformation matrix is ​​a transformation matrix obtained in advance through camera calibration. This matrix can convert the pixel coordinates of the image plane into planar coordinates from a bird's-eye view, restoring the true spatial proportions of the scene. The lane line bird's-eye view coordinate set is the coordinate data obtained after the lane line pixel coordinates have undergone inverse perspective transformation. This coordinate set eliminates perspective distortion and can realistically reflect the actual distribution and geometric shape of the lane lines on the road surface. Polynomial fitting is a processing method that uses polynomial functions to approximate discrete coordinate points using curves.

[0091] By fitting, a smooth and continuous curve expression can be obtained to describe the overall geometric pattern of the lane lines. The 3D geometric model of the lane lines is a spatial geometric model constructed based on the fitting results of the bird's-eye view coordinates. This model can reflect the spatial direction and positional relationship of the lane lines, providing geometric constraints for spatial positioning. The bounding box coordinates in the station number information are the coordinates of the outer rectangle of the detected station number in the image. These coordinates can determine the specific location and area occupied by the station number in the image. The region of interest is a local area related to the target selected from the entire image. A binocular image pair is a pair of images captured simultaneously by the two lenses of a binocular camera. The two images have disparity differences, providing basic data for stereo matching and depth calculation. Regionalized stereo matching is a method of pixel matching of binocular images within a specified local area. Through matching, the positional differences between corresponding pixels can be found, i.e., disparity information. The station number area disparity map is image data that records the disparity value of each pixel within the station number area. The value at each position in the image represents the disparity magnitude of that point in the binocular image. Valid disparity pixels are pixels with reliable disparity values ​​that meet preset conditions. These pixels accurately reflect the spatial location of the target, eliminating interference data caused by incorrect matching. The binocular baseline distance is the actual distance between the optical centers of the two cameras in a binocular camera system. This distance is a fixed physical quantity and directly participates in the parallax-to-depth conversion calculation. The preset distance threshold is a pre-defined reasonable range boundary for depth values. Invalid values ​​refer to depth values ​​that exceed the reasonable range or contain obvious errors. The first depth value is the lane line geometric constraint depth value after reasonableness verification. The second depth value is the binocular parallax depth value after reasonableness verification.

[0092] Specifically, firstly, the set of lane line pixel coordinates is input into a pre-calibrated inverse perspective transformation matrix. This matrix is ​​constructed based on the planar assumption after obtaining the camera's intrinsic and extrinsic parameters through offline calibration. It maps the pixel coordinates in the image coordinate system to the bird's-eye view coordinate system, obtaining the lane line bird's-eye view coordinate set. This eliminates perspective distortion, ensuring that parallel lane lines maintain equidistant parallelism in the bird's-eye view. Secondly, the lane line bird's-eye view coordinate set is fitted with a third-order polynomial using the least squares method to construct a three-dimensional geometric model of the lane lines and calculate their three-dimensional positions in the world coordinate system. Combining the pixel coordinates of the bottom midpoint of the bounding box in the station information, the corresponding longitudinal distance is calculated through interpolation using the lane line geometric model. The lateral distance is taken as the lane half-width plus the offset of the bounding box center from the lane centerline. The initial first depth value is calculated based on the Pythagorean theorem, and stable relative distance estimates are provided using lane line structural constraints. Finally, the bounding box coordinates of the station are extracted and extended outward by 10% as the region of interest. The RAFT-Stereo stereo matching network is used to input the ROI region of the binocular image pair. A dense disparity map is obtained through 4D cost volume correlation calculation and iterative optical flow update. The disparity values ​​are screened for effective pixels by Sobel gradient consistency test and left-right consistency test. After removing outliers in occluded areas and texture-deficient areas, the effective disparity mean is calculated. Combined with the camera focal length and binocular baseline distance, the initial second depth value is calculated based on the triangulation principle. Stereo vision is used to provide accurate absolute distance estimation. Finally, the initial first depth value and the initial second depth value are compared with the preset distance thresholds [5m, 200m] respectively. Outliers that exceed the effective ranging range are removed. The depth values ​​that pass the verification are checked for temporal consistency. If the depth difference with the previous frame exceeds 10m, it is marked as a suspicious value and the weight is reduced. Finally, the first and second depth values ​​that have been double-verified are output, providing reliable dual-source depth data for subsequent adaptive weighted fusion.

[0093] Step S50: Perform adaptive weighted fusion based on the first depth value and the second depth value to obtain the target depth value.

[0094] It should be noted that step S50 includes: extracting the lane line detection confidence value corresponding to the first depth value and the marker detection confidence value corresponding to the second depth value to obtain two sets of detection confidence value values; calculating the historical ranging error variance of the first depth value and the second depth value within a preset time window to obtain two sets of error variance data; inputting the two sets of detection confidence value values ​​and the two sets of error variance data into the weight calculation model, and obtaining the first weight of the first depth value and the second weight of the second depth value through normalization processing; multiplying the first depth value and the first weight to obtain the first weighted depth value, and multiplying the second depth value and the second weight to obtain the second weighted depth value; summing the first weighted depth value and the second weighted depth value to obtain the initial fusion depth value; and performing neighborhood smoothing filtering on the initial fusion depth value to obtain the target depth value.

[0095] It's important to understand that lane line detection confidence is a numerical value that measures the reliability of the detection results for the set of lane line pixel coordinates. This value reflects the accuracy of lane line feature extraction; a higher value indicates a more reliable detection result. Pile marker detection confidence is a numerical value that measures the reliability of the detection results for pile marker information. This value reflects the accuracy of pile marker target detection and text recognition; a higher value indicates a more reliable detection result. The two sets of detection confidence values ​​are combinations of confidence data corresponding to the first and second depth values, respectively. One set represents lane line detection confidence, and the other represents pile marker detection confidence. The preset time window is a pre-defined continuous time interval. This interval is used to statistically analyze historical ranging data within a specified time period, ensuring that the error variance calculation reflects recent ranging stability. The historical ranging error variance is the average of the squares of the deviations between the depth measurement values ​​and the true values ​​within the preset time window. This value reflects the stability of the depth ranging results; a smaller variance indicates a more stable ranging result, and higher priority is given in weight allocation. The weight calculation model is an algorithmic model used to calculate the weights of each depth value based on the input parameters. The model takes detection confidence and error variance as inputs and uses pre-defined logic to allocate weights reasonably, ensuring that the weights are positively correlated with the reliability of the depth values. Neighborhood smoothing filtering is a method of smoothing local regions of the initial fused depth values. By calculating the mean or weighted mean of the depth values ​​of the target pixel and its surrounding pixels, isolated outliers are eliminated, resulting in a smoother depth result.

[0096] Specifically, firstly, the lane detection confidence corresponding to the first depth value is extracted. This confidence is calculated from the average activation value of the Sigmoid output probability map of the lane segmentation head in the instance region. Simultaneously, the marker detection confidence corresponding to the second depth value is extracted. This confidence is obtained by multiplying the target existence probability of the YOLO detection head by the category confidence. The two sets of values ​​are normalized to the [0,1] interval to form a detection confidence value set. Secondly, a sliding time window of 30 frames is maintained, and the residual sequences of the first and second depth values ​​relative to the reference true value or time-consistent estimate are recorded within this window. The sample variance of the two sets of residuals is calculated as the historical ranging error variance data. The larger the variance, the worse the recent stability of the method. Then... The detection confidence score and error variance data are input into the weight calculation model to calculate the first weight and the second weight, so that depth values ​​with high confidence and small variance are given greater weight. Softmax normalization is used to ensure that the sum of the weights is 1. Then, the first depth value is multiplied by the first weight to obtain the first weighted depth value, and the second depth value is multiplied by the second weight to obtain the second weighted depth value. The sum of the two is used to obtain the initial fusion depth value, realizing the complementary advantages of geometric constraints and stereo vision. Finally, the initial fusion depth value is subjected to a 5-frame neighborhood smoothing filter, and a Gaussian weight kernel function is used to weight the fusion results of the current frame and the two frames before and after it to suppress random jitter in a single frame and output a temporally coherent target depth value, taking into account both real-time response and estimation stability.

[0097] Step S60: Perform time-consistent spatiotemporal fusion and correlation reasoning on the target depth value to obtain a smooth depth estimate and output semantic location information.

[0098] It should be noted that step S60 includes: performing multi-target tracking on the set of lane line pixel coordinates, station mark information, and corresponding target depth values ​​of multiple consecutive frames, assigning unique identifiers across frames to obtain an identifiable spatiotemporal perception dataset; calculating the lateral distance between the station mark position corresponding to the station mark information and the set of pixel coordinates of each lane line, establishing a spatiotemporal association pair between the station mark and its corresponding lane line, and obtaining an associated depth dataset; extracting the perception quality score corresponding to the environmental perception data, adjusting the process noise covariance of the adaptive Kalman filter according to the perception quality score, and obtaining a parameter-calibrated Kalman filter; inputting the target depth value in the associated depth dataset into the parameter-calibrated Kalman filter, performing temporal smoothing filtering to obtain a smoothed depth estimate; integrating the smoothed depth estimate, the station mark text in the station mark information, the timestamp of image acquisition, and the environmental perception data to generate a structured location data set; and standardizing the format and labeling the confidence level of the structured location data set to obtain semantic location information containing station mark information, depth distance, timestamp, and environmental state.

[0099] It's important to understand that multi-target tracking is a continuous tracking operation involving multiple targets within an image sequence. A unique cross-frame identifier is a unique code assigned to each tracked target. This code remains unique and fixed across all consecutive frames, accurately distinguishing different targets and enabling the integration of cross-frame information for the same target. The marker location represents the specific position of the marker in image space. Lateral distance is the vertical distance from the marker location to the lane lines represented by the set of pixel coordinates of each lane line. This distance is the core criterion for determining the lane to which the marker belongs; the lane line corresponding to the smallest lateral distance is the lane to which the marker belongs. Spatiotemporal association pairs are the one-to-one correspondence between markers and their respective lane lines. The perception quality score is a quantitative evaluation of the perception effect based on environmental perception data. A higher score indicates a more favorable environment for feature extraction and depth calculation of lane lines and markers, resulting in stronger reliability of the perception results. An adaptive Kalman filter is a filtering algorithm tool that can dynamically adjust parameters based on external input. Semantic location information is the final output containing multi-dimensional semantic location description information. Semantic location information includes station information, depth distance, timestamp, and environmental status.

[0100] Specifically, Kalman filtering is used to track the lane line pixel coordinates of multiple consecutive frames to maintain lane line ID consistency. For station marker information, the DeepSORT algorithm is used in conjunction with bounding box IoU matching and appearance feature cosine similarity calculation for cross-frame association. A unique identifier, TrackID, is assigned to each lane line instance and station marker target that has been stably tracked for more than 3 frames, resulting in an identifiable spatiotemporal perception dataset, thus resolving ID jump issues caused by occlusion or brief missed detections. Secondly, the bottom center point of the station marker bounding box is projected onto the bird's-eye view space, and the lateral Euclidean distance between this point and the 3D geometric models of each lane line at the same longitudinal distance is calculated. The lane line with the smallest distance, less than the lane half-width threshold, is selected as the lane to which it belongs, establishing a "station marker Track ID - lane line Track" relationship. The spatial-temporal association pair of "ID" is used to bind the target depth value to the association relationship to form an associated depth dataset, ensuring semantic consistency between depth estimation and lane-level localization. Then, the image quality index and weather condition score in the environment coding vector are extracted and weighted to calculate the perception quality score Q. When Q is below 0.3, it is judged as a high uncertainty scene. The diagonal elements of the process noise covariance matrix of the Kalman filter are increased to 3 times the default value to make the filter more trusting of the current observation and rapid response to changes. When Q is above 0.7, it is decreased to 0.5 times the default value to make the filter more dependent on the smooth estimation output of the motion model, resulting in a parameter-calibrated Kalman filter, realizing the temporal fusion intensity adjustment of environment-adaptive. Subsequently, the target depth value of each station mark target in the associated depth dataset is used as the observation value and input into the Kalman filter of the corresponding TrackID. After the prediction update step, the posterior state estimate is output as the smooth depth estimate value, and historical trajectory information is used to suppress single-frame perception noise. Finally, the smooth depth estimate value, station text, image timestamp, and environment state code are encapsulated in a preset JSON format. The depth value is retained to two decimal places, the station text is verified to the national standard format, and the timestamp is converted to ISO. The system follows the 8601 standard and calculates a comprehensive confidence level based on tracking stability and depth estimation variance. It then generates semantic location information, including station information, depth distance, timestamp, and environmental status, which is output to the vehicle terminal or cloud database to support navigation positioning and road maintenance decisions.

[0101] This embodiment acquires and preprocesses road images using a binocular camera. The preprocessed images are then input into a multi-task fusion perception network containing a shared encoder and temporal convolutional units to extract lane line pixel coordinates and station information. Dual depth values ​​are obtained through multi-constraint joint ranging, followed by adaptive weighted fusion to obtain the target depth value. Finally, temporal fusion inference yields a smooth depth estimate, outputting semantic location information including station number and distance. By fusing lane line structure and station semantic information for ranging, no additional sensors are required, resulting in low cost and easy deployment. This improves the accuracy and robustness of road station depth ranging in complex environments.

[0102] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The depth ranging method S30, which integrates road lane lines and marker posts, further includes steps S201 to S205:

[0103] Step S201: The preprocessed image of the current processing frame and the preprocessed images of the previous preset number of frames are combined to form a temporal image sequence, which is then input into the shared encoder of the preset multi-task fusion perception network to obtain multi-scale spatial features.

[0104] It should be noted that the current processing frame refers to the single preprocessed image in which feature extraction and analysis are being performed. The preset number of preceding frames refers to a fixed number of preprocessed image frames acquired consecutively before the current processing frame. This number of frames is a fixed value set in advance.

[0105] Specifically, five preprocessed images—the current frame and its four preceding frames—are extracted from the circular buffer in reverse timestamp order. The 512×512×3 RGB tensors of each frame are then stacked along the batch dimension to form a four-dimensional temporal image sequence of shape (5, 3, 512, 512). The first dimension, 5, represents the five-frame temporal sequence. The second dimension, 3, represents the RGB three channels. The third dimension, 512, represents the image height. The fourth dimension, 512, represents the image width. Optical flow fields are calculated using inter-frame pixel-level differences as motion compensation references to ensure spatiotemporal alignment of features on the same road in dynamic scenes, providing spatial correspondence for subsequent temporal fusion. Secondly, the temporal image sequence is input into a shared encoder of a pre-defined multi-task fusion perception network. This shared encoder employs a multi-stage residual structure. The first stage uses a 7×7 convolution with a stride of 2 to downsample to 256×256, followed by three bottleneck residual blocks to extract edge and color features. Each bottleneck residual block consists of a 1×1 convolution to reduce dimensionality to 64 channels, a 3×3 spatial convolution, and a 1×1 convolution to increase dimensionality to 256 channels. The second to fourth stages use 1×1 convolutions with a stride of 2 to achieve downsampling followed by stacking 4, 6, and 3 bottleneck residual blocks, respectively, with the number of channels progressively expanding to 512 and 1024. 2048, forming a multi-scale spatial feature pyramid with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image. The bottleneck structure reduces the computational cost of 3×3 convolution by first reducing and then increasing the dimensionality. Residual skip connections alleviate the vanishing gradient in deep layers, enabling the network to extract multi-level features from lane line details to road scene semantics under limited computing power. Finally, the multi-scale spatial features output from the four stages are stored separately according to resolution. High-resolution low-level features retain lane line edge texture information for accurate localization, while low-resolution high-dimensional features encode the station number and the overall semantic context of the road for robust recognition. This provides a hierarchical feature foundation for the subsequent environment-adaptive fusion of temporal convolutional units and the specific processing of dual-task decoders, balancing the requirements for small target detection accuracy and structural line extraction stability.

[0106] Step S202: Temporally fuse multi-scale spatial features using temporal convolutional units to obtain temporally enhanced features.

[0107] Specifically, firstly, the multi-scale spatial features output from each stage of the shared encoder are input into the corresponding temporal convolutional units according to each stage. That is, a ConvLSTM unit is embedded at the output of each of the four stages: conv2_x, conv3_x, conv4_x, and conv5_x. Each unit receives the spatial feature map of that stage in the current frame as the input for the current time step, and simultaneously receives the hidden state and cell state from the previous time step. Through a triple gating mechanism of input gate, forget gate, and output gate, the candidate state at the current time step is calculated, the cell memory is updated, and a new hidden state is output, realizing the dynamic aggregation of spatial features in the temporal dimension. Secondly, for each stage of the ConvLSTM... The vLSTM output is weighted by channel attention. The final hidden state of the 5-frame temporal sequence is concatenated along the channel dimension and compressed to the original number of channels by 1×1 convolution. Then, it is added to the residual of the current frame feature to enhance the contribution of recent frames to the current estimate while suppressing the outdated information of distant frames, resulting in a temporal enhancement feature that integrates historical motion consistency and current observation freshness. Finally, the temporal enhancement features of the four stages are output at the original scale to form a temporal enhancement feature pyramid at the same level as the multi-scale spatial feature pyramid. This enables lane line features to have cross-frame continuity constraints and station sign features to have motion trajectory smoothness priors, improving the ability to recover from occlusion and compensate for missed detections in dynamic scenes.

[0108] Step S203: Input the temporal enhancement features into the environmental condition modulation module, generate modulation parameters for lane line branches and station mark branches based on the environmental coding vector, and obtain the modulated features.

[0109] Specifically, firstly, the highest-resolution 1 / 4-scale feature map from the temporal enhancement features output by the shared encoder is used as the input to the environmental condition modulation module. Simultaneously, the environmental encoding vector corresponding to the current frame is extracted from the synchronization buffer. This vector contains three quantization dimensions: light intensity level, fog concentration score, and rain / snow score. Secondly, the environmental encoding vector is input into a small MLP network consisting of two fully connected layers. The first layer expands the 3D environmental features to a 64-dimensional intermediate representation and introduces nonlinearity through ReLU activation. The second layer outputs two sets of modulation parameters in parallel: the first scaling factor and first bias parameter for lane line branches, and the second scaling factor and second bias parameter for station marker branches. Each set of parameters has the same dimensionality as the input feature channels. An end-to-end training loop is then used to learn the modulation. The context-task mapping relationship is established. Then, the temporal enhancement features are fed into the batch normalization layer of the dual-branch system for channel normalization to obtain intermediate features with zero mean and unit variance. Channel-by-channel affine transformation is performed on the lane line branch and the station sign branch. The second scaling factor is automatically increased in foggy scenes to enhance edge contrast and automatically decreased in rainy scenes to suppress noise amplification. The second bias parameter adjusts the feature activation baseline according to the lighting conditions. Finally, the modulated dual-branch features are output to the subsequent decoder, enabling the lane line branch to enhance the lane line edge gradient response in low visibility environments and the station sign branch to suppress high-frequency rain and fog interference in severe weather. This achieves environmental adaptive dynamic calibration of the feature representation and improves the task-specific perception capability in complex scenes.

[0110] Step S204: Input the modulated features into the lane line branch of the multi-task decoder, perform lane line instance segmentation and curve fitting on the modulated features, and obtain the set of pixel coordinates of the lane lines.

[0111] Specifically, the modulated features of the lane branch are first input into a U-shaped structure of upsampling-convolution-upsampling. The first-level transposed convolution upsamples the 1 / 32 scale features to a 1 / 16 scale and connects them to the corresponding layer features of the encoder via a skip connection. Then, a 3×3 convolution fuses semantic and detail information. The second-level transposed convolution upsamples the features to a 1 / 8 scale, and the process is repeated. Finally, the third-level transposed convolution restores the original features. Figure 1The process involves: 1) Using a resolution of 2 / 2, a high-resolution feature map is generated to preserve lane line edge details. 2) Inputting the high-resolution feature map into a segmentation head consisting of a 1×1 convolution and a Sigmoid activation function, the output is a pixel-level lane line existence probability map. A probability threshold of 0.5 is set for binarization to obtain an initial segmentation mask. The DBSCAN density clustering algorithm is used to cluster connected regions in the mask according to spatial distance, distinguishing different lane line instances and filtering out noise regions with an area less than 100 pixels, extracting the pixel coordinate set of each instance. Then, the pixel coordinate set is input into a pre-calibrated inverse perspective transformation matrix and mapped to the bird's-eye view space to eliminate perspective distortion and make the lane lines parallel and equidistant. The bird's-eye coordinates are fitted with a third-order polynomial curve using the least squares method, iteratively removing outliers with residuals greater than 3 times the standard deviation to optimize the fitting accuracy, resulting in a smooth lane line geometric model. Finally, the fitted curve is sampled at equal intervals longitudinally and reprojected into the image space, outputting the pixel coordinate set of each lane line, providing a road structure constraint benchmark for subsequent lateral positioning and depth calculation of station markers.

[0112] Step S205: Through the station number branch of the multi-task decoder, target detection and text recognition are performed on the modulated features to obtain station number information containing bounding box coordinates and station number text.

[0113] Specifically, firstly, the modulated features of the branch road of the station number are input into the feature pyramid fusion structure. The modulated features at three scales (1 / 32, 1 / 16, and 1 / 8) are aligned to 256 channels through 1×1 convolutions. From top to bottom, they are upsampled layer by layer through horizontal connections and added element-wise to the features of the lower layer. Finally, 3×3 convolutions are added at the output of each scale to eliminate aliasing, forming an enhanced feature pyramid that integrates multi-scale semantics, accommodating both distant small targets and nearby large targets. Secondly, the enhanced feature pyramid is input into an improved YOLO detection head. Three aspect ratio anchor boxes are preset on each scale grid cell. Parallel convolutions predict the bounding box offset, target presence confidence, and class probability. Non-maximum suppression is used to filter overlapping boxes with a confidence threshold of 0.6. The output is the coordinates of the station number bounding box and its detection confidence. The bounding box is expanded outward by 10% and cropped and normalized to a fixed-size 224×64 text region image. Then, the text region image is input into an enhanced lightweight PaddleOCR. First, the character region is located through a DB text detection network with multi-scale feature enhancement. Then, a micro-network with residual block embedding and asymmetric convolution decomposition is used for text correction and feature extraction. The standard convolution is replaced with a Ghost module to reduce computational complexity. Then, a simplified GRU and single-head self-attention hybrid module is used for sequence modeling to output the station number text character sequence. Finally, the recognition results are post-processed and verified to filter out abnormal text in non-standard station number formats. The bounding box coordinates are bound to the verified station number text to form a station number information output containing spatial location and semantic content, providing absolute mileage identification for subsequent depth ranging and location association.

[0114] In this embodiment, the current frame and the preprocessed images of the previous preset number of frames are combined to form a temporal image sequence. This sequence is then input into the shared encoder of a preset multi-task fusion perception network to obtain multi-scale spatial features. Temporal enhanced features are obtained through temporal fusion by a temporal convolution unit and input into an environmental condition modulation module to generate modulation parameters and obtain modulated features. Finally, through the dual branches of the multi-task decoder, lane line instance segmentation and curve fitting, station mark target detection and text recognition are completed respectively, resulting in the set of lane line pixel coordinates and station mark information. This enables synchronous perception of lane lines and station marks, and the dual-branch targeted detection improves accuracy, providing high-quality feature support for subsequent depth ranging.

[0115] Based on the first embodiment of this application, this application also provides a depth ranging device that integrates road lane lines and marker posts. Please refer to... Figure 3 The device includes:

[0116] The acquisition module 10 is used to acquire road image sequences through a binocular camera.

[0117] The preprocessing module 20 is used to preprocess the road image sequence to obtain a preprocessed image.

[0118] The fusion perception module 30 is used to process the preprocessed image input to a preset multi-task fusion perception network to obtain the pixel coordinate set of lane lines and station number information. The preset multi-task fusion perception network includes a shared encoder, a temporal convolutional unit, an environmental condition modulation module, and a multi-task decoder. The shared encoder adopts a multi-stage residual structure and embeds a temporal convolutional unit at the output end. The multi-task decoder adopts a dual-branch parallel structure, which includes a lane line branch and a station number branch. The lane line branch adopts an upsampling-convolution-upsampling U-shaped structure, and the station number branch adopts a feature pyramid fusion structure. The environmental condition modulation module consists of a fully connected network and a batch normalization layer.

[0119] The joint ranging module 40 is used to calculate the first depth value and the second depth value based on the pixel coordinate set of the lane line and the station information through a multi-constraint joint depth ranging algorithm.

[0120] The weighted calculation module 50 is used to perform adaptive weighted fusion based on the first depth value and the second depth value to obtain the target depth value.

[0121] The results module 60 is used to perform time-consistent spatiotemporal fusion and correlation reasoning on the target depth value, obtain a smooth depth estimate and output semantic location information, which includes station information, depth distance, timestamp and environmental state.

[0122] The depth ranging device integrating road lane lines and mileage markers provided in this application, employing the depth ranging method integrating road lane lines and mileage markers described in the above embodiments, can solve the technical problem of how to improve the accuracy and robustness of road mileage marker depth ranging in complex environments. Compared with the prior art, the beneficial effects of the depth ranging device integrating road lane lines and mileage markers provided in this application are the same as those of the depth ranging method integrating road lane lines and mileage markers provided in the above embodiments, and other technical features in the depth ranging device integrating road lane lines and mileage markers are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0123] This application provides a depth ranging device that integrates road lane lines and station markers. The depth ranging device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the depth ranging method integrating road lane lines and station markers in the above embodiment 1.

[0124] The following is for reference. Figure 4This document illustrates a structural schematic diagram of a depth ranging device suitable for implementing the embodiments of this application that integrates road lane lines and station markers. The depth ranging device integrating road lane lines and station markers in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The depth ranging device shown that integrates road lane lines and station markers is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0125] like Figure 4 As shown, the depth ranging device integrating road lane lines and station markers may include a processing unit 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in ROM 1002 (Read Only Memory) or a program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the depth ranging device integrating road lane lines and station markers. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the depth ranging device integrating road lane lines and station markers to communicate wirelessly or wiredly with other devices to exchange data. Although various depth ranging devices integrating road lane lines and station markers are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer may be implemented alternatively.

[0126] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0127] The depth ranging device integrating road lane lines and mileage markers provided in this application, employing the depth ranging method integrating road lane lines and mileage markers described in the above embodiments, can solve the technical problem of how to improve the accuracy and robustness of road mileage marker depth ranging in complex environments. Compared with the prior art, the beneficial effects of the depth ranging device integrating road lane lines and mileage markers provided in this application are the same as those of the depth ranging method integrating road lane lines and mileage markers provided in the above embodiments, and other technical features in this depth ranging device integrating road lane lines and mileage markers are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0128] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0129] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0130] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the depth ranging method for fusing road lane lines and station markers in the above embodiments.

[0131] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0132] The aforementioned computer-readable medium may be included in a depth ranging device that integrates road lane lines and station markers; or it may exist independently and not be assembled into a depth ranging device that integrates road lane lines and station markers.

[0133] The aforementioned computer-readable medium carries one or more programs that, when executed by a depth ranging device integrating road lane lines and station markers, enable the depth ranging device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0134] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0135] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0136] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned depth ranging method that integrates road lane lines and station markers. This addresses the technical problem of improving the accuracy and robustness of road station depth ranging in complex environments. Compared to the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the depth ranging method integrating road lane lines and station markers provided in the above embodiments, and will not be elaborated upon here.

[0137] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the depth ranging method for fusing road lane lines and station markers as described above.

[0138] The computer program product provided in this application can solve the technical problem of how to improve the accuracy and robustness of road station depth measurement in complex environments. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the depth measurement method that integrates road lane lines and station signs provided in the above embodiments, and will not be repeated here.

[0139] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A depth ranging method integrating road lane lines and marker posts, characterized in that, The method includes: Acquire road image sequences using a binocular camera; The road image sequence is preprocessed to obtain a preprocessed image; The preprocessed image is input into a preset multi-task fusion perception network for processing to obtain the pixel coordinate set of lane lines and station number information. The preset multi-task fusion perception network includes a shared encoder, a temporal convolutional unit, an environmental condition modulation module, and a multi-task decoder. The shared encoder adopts a multi-stage residual structure and embeds a temporal convolutional unit at the output end. The multi-task decoder adopts a dual-branch parallel structure, which includes a lane line branch and a station number branch. The lane line branch adopts an upsampling-convolution-upsampling U-shaped structure, and the station number branch adopts a feature pyramid fusion structure. The environmental condition modulation module consists of a fully connected network and a batch normalization layer. Based on the set of pixel coordinates of the lane lines and the information of the station markers, a first depth value and a second depth value are calculated using a multi-constraint joint depth ranging algorithm. The target depth value is obtained by adaptive weighted fusion of the first depth value and the second depth value. The target depth value is subjected to time-consistent spatiotemporal fusion and correlation reasoning to obtain a smooth depth estimate and output semantic location information, wherein the semantic location information includes station information, depth distance, timestamp and environmental state; The step of calculating the first depth value and the second depth value based on the pixel coordinate set of the lane line and the station marker information using a multi-constraint joint depth ranging algorithm includes: Input the set of pixel coordinates of the lane lines into a pre-calibrated inverse perspective transformation matrix to complete the conversion from pixel coordinates to bird's-eye view coordinates, and obtain the set of bird's-eye view coordinates of the lane lines. A three-dimensional geometric model of the lane lines is constructed by performing polynomial fitting on the bird's-eye view coordinate set of the lane lines. Based on the three-dimensional geometric model of the lane line and the bounding box coordinates in the station sign information, the world coordinate system coordinates of the bottom midpoint of the station sign are determined to obtain the initial first depth value. The bounding box coordinates in the station number information are extracted as the region of interest. The binocular image pairs acquired by the binocular camera are subjected to regional stereo matching to obtain the disparity map of the station number region. The effective disparity pixels of the disparity map of the mileage marker area are filtered and the mean is calculated to obtain the effective disparity value of the mileage marker area. The effective parallax value of the station marker area is combined with the camera intrinsic parameters and the binocular baseline distance to calculate the initial second depth value; The initial first depth value and the initial second depth value are respectively checked for reasonableness, and invalid values ​​that exceed the preset distance threshold are removed to obtain the first depth value and the second depth value.

2. The depth ranging method integrating road lane lines and station markers as described in claim 1, characterized in that, The step of preprocessing the road image sequence to obtain a preprocessed image includes: Collect environmental perception data, perform timestamp alignment and standardization processing on the environmental perception data, and generate an environmental coding vector; Based on the environmental encoding vector, an adaptive adjustment instruction for exposure parameters is generated; Based on the adaptive adjustment command for the exposure parameters, the exposure time and gain of the binocular camera are dynamically adjusted to obtain a road image sequence with adjusted parameters. Based on the weather score in the environmental coding vector, environmentally-aware image enhancement processing is performed on the road image sequence after parameter adjustment to obtain a preprocessed image.

3. The depth ranging method integrating road lane lines and station markers as described in claim 2, characterized in that, The step of performing environment-aware image enhancement processing on the road image sequence adjusted according to the weather score in the environment coding vector to obtain a preprocessed image includes: Obtain a weather score, which includes a fog concentration score, a rain / snow score, and a visibility score; When the fog concentration score of the weather rating is greater than the preset fog concentration threshold, a lightweight defogging algorithm is enabled on the road image sequence after parameter adjustment to obtain the first preprocessed image; When the rain and snow score of the weather score is greater than the preset rain and snow threshold, a preset rain pattern suppression filter is applied to the road image sequence after parameter adjustment to obtain a second preprocessed image. When the visibility score of the weather rating is lower than the preset visibility threshold, the ROI region of the road image sequence after parameter adjustment is reduced to obtain a third preprocessed image; When the fog concentration score of the weather score is greater than the preset fog concentration threshold, and the rain and snow score of the weather score is greater than the preset rain and snow threshold, the lightweight defogging algorithm and the preset rain pattern suppression filter are sequentially applied to the road image sequence after parameter adjustment to obtain the fourth preprocessed image. When the fog concentration score of the weather rating is greater than the preset fog concentration threshold and the visibility score of the weather rating is lower than the preset visibility threshold, a lightweight defogging algorithm and a reduction of the ROI region are sequentially applied to the road image sequence after parameter adjustment to obtain the fifth preprocessed image. When the rain / snow score of the weather rating is greater than the preset rain / snow threshold and the visibility score of the weather rating is lower than the preset visibility threshold, the preset rain pattern suppression filter and the ROI region are applied sequentially to the road image sequence after parameter adjustment to obtain the sixth preprocessed image. When the fog concentration score of the weather score is greater than the preset fog concentration threshold, the rain and snow score of the weather score is greater than the preset rain and snow threshold, and the visibility score of the weather score is lower than the preset visibility threshold, the lightweight defogging algorithm, the preset rain pattern suppression filter, and the ROI region are sequentially applied to the road image sequence after parameter adjustment to obtain the seventh preprocessed image. Select the corresponding image from the first preprocessed image, the second preprocessed image, the third preprocessed image, the fourth preprocessed image, the fifth preprocessed image, the sixth preprocessed image, and the seventh preprocessed image as the preprocessed image.

4. The depth ranging method integrating road lane lines and station markers as described in claim 1, characterized in that, The step of processing the preprocessed image into a preset multi-task fusion perception network to obtain the pixel coordinate set of lane lines and marker information includes: The preprocessed image of the current processing frame and the preprocessed images of the previous preset number of frames are combined to form a temporal image sequence, which is then input into the shared encoder of the preset multi-task fusion perception network to obtain multi-scale spatial features. Temporal fusion of the multi-scale spatial features is performed using the temporal convolutional unit to obtain temporally enhanced features; The temporal enhancement features are input into the environmental condition modulation module, and the modulation parameters of lane line branches and station mark branches are generated according to the environmental coding vector to obtain the modulated features. The modulated features are input into the lane line branch of the multi-task decoder, and lane line instance segmentation and curve fitting are performed on the modulated features to obtain the set of pixel coordinates of the lane lines. The post number branch of the multi-task decoder is used to perform post number target detection and text recognition on the modulated features to obtain post number information containing bounding box coordinates and post number text.

5. The depth ranging method integrating road lane lines and station markers as described in claim 1, characterized in that, The step of adaptively weighting and fusing the first depth value and the second depth value to obtain the target depth value includes: Extract the lane line detection confidence score corresponding to the first depth value and the station marker detection confidence score corresponding to the second depth value to obtain two sets of detection confidence scores. The historical ranging error variances of the first depth value and the second depth value within a preset time window are statistically analyzed to obtain two sets of error variance data. The two sets of detection confidence values ​​and the two sets of error variance data are input into the weight calculation model, and the first weight of the first depth value and the second weight of the second depth value are obtained through normalization processing. Multiply the first depth value by the first weight to obtain the first weighted depth value, and multiply the second depth value by the second weight to obtain the second weighted depth value. The first weighted depth value and the second weighted depth value are summed to obtain the initial fusion depth value; The initial fusion depth value is subjected to neighborhood smoothing filtering to obtain the target depth value.

6. The depth ranging method integrating road lane lines and station markers as described in claim 1, characterized in that, The step of performing time-consistent spatiotemporal fusion and correlation reasoning on the target depth value to obtain a smooth depth estimate and output semantic location information includes: Multi-target tracking is performed on the set of lane line pixel coordinates, station number information and corresponding target depth values ​​of multiple consecutive frames, and a unique identifier is assigned across frames to obtain an identifiable spatiotemporal perception dataset. Calculate the lateral distance between the location of the mileage marker corresponding to the mileage marker information and the pixel coordinate set of each lane line, establish a spatiotemporal association pair between the mileage marker and its corresponding lane line, and obtain the associated depth dataset; Extract the perception quality score corresponding to the environmental perception data, and adjust the process noise covariance of the adaptive Kalman filter according to the perception quality score to obtain the parameter-calibrated Kalman filter. The target depth value in the associated depth dataset is input into the parameter-calibrated Kalman filter for time-series smoothing filtering to obtain a smoothed depth estimate. By integrating the smoothed depth estimate, the station number text in the station number information, the timestamp of image acquisition, and the environmental perception data, a structured location data set is generated; The structured location data set is formatted and labeled with confidence to obtain semantic location information containing station information, depth distance, timestamp and environmental status.

7. A depth ranging device integrating road lane lines and marker posts, characterized in that, The device is applied to the depth ranging method that integrates road lane lines and station markers as described in any one of claims 1-6, and the device comprises: The acquisition module is used to acquire road image sequences using a binocular camera; The preprocessing module is used to preprocess the road image sequence to obtain a preprocessed image; The fusion perception module is used to process the preprocessed image input into a preset multi-task fusion perception network to obtain the pixel coordinate set of lane lines and station number information. The preset multi-task fusion perception network includes a shared encoder, a temporal convolutional unit, an environmental condition modulation module, and a multi-task decoder. The shared encoder adopts a multi-stage residual structure and embeds a temporal convolutional unit at the output end. The multi-task decoder adopts a dual-branch parallel structure, which includes a lane line branch and a station number branch. The lane line branch adopts an upsampling-convolution-upsampling U-shaped structure, and the station number branch adopts a feature pyramid fusion structure. The environmental condition modulation module consists of a fully connected network and a batch normalization layer. The joint ranging module is used to calculate, based on the set of pixel coordinates of the lane lines and the station marker information, a multi-constraint joint depth ranging algorithm to obtain a first depth value and a second depth value. The weighted calculation module is used to perform adaptive weighted fusion based on the first depth value and the second depth value to obtain the target depth value; The results module is used to perform time-consistent spatiotemporal fusion and correlation reasoning on the target depth value to obtain a smooth depth estimate and output semantic location information, wherein the semantic location information includes station information, depth distance, timestamp and environmental state.

8. A depth ranging device integrating road lane lines and marker posts, characterized in that, The device includes: a memory, a processor, and a depth ranging program for fusing road lane lines and station markers stored in the memory and running on the processor, the depth ranging program for fusing road lane lines and station markers being configured to implement the steps of the depth ranging method for fusing road lane lines and station markers as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a depth ranging program that integrates road lane lines and station markers. When the processor executes the depth ranging program that integrates road lane lines and station markers, it implements the steps of the depth ranging method for integrating road lane lines and station markers as described in any one of claims 1-6.

Images