An image feature calibration method based on motion offset

Abstract

This invention discloses an image feature calibration method based on motion offset, relating to the field of track inspection image processing. The method includes steps of motion image acquisition, initial image calibration, fuzzy feature calibration, and calibrated image output. Real-time acceleration data is acquired using a train acceleration sensor and converted into pixel displacement deviation for initial calibration to obtain a reference image. Based on target features from a historical database, the target subject and environmental region are separated, and an environmental point processing curve is generated by combining motion parameters. A set of pixels to be calibrated is selected through feature fitting. A fuzzy feature calibration strategy is used to calibrate the pixels, which are then fused with the reference image. The calibration accuracy is dynamically optimized through curve calibration and effect verification steps. This invention achieves layered calibration, improves the detail reproduction of motion images, enhances environmental adaptability, and ensures the integrity of target features, making it suitable for rail transit image monitoring scenarios.

Description

Technical Field

[0001] This invention relates to the field of track detection image processing, and more specifically to an image feature calibration method based on motion offset. Background Technology

[0002] With the rapid development of intelligent rail transit, image acquisition equipment on rail trains has been widely used in key scenarios such as track environment monitoring, line inspection, and safety early warning. However, during high-speed operation, the image acquisition equipment will experience periodic or sudden motion shifts due to factors such as acceleration during startup, deceleration during braking, and centrifugal force on track curves. This results in quality problems such as pixel displacement, edge blurring, and motion ghosting in the acquired moving images.

[0003] In existing technologies, calibration methods for moving images are mainly divided into two categories: one is based on pure vision algorithms, which achieves offset correction through image feature point matching. However, this type of method is highly dependent on image texture features and is prone to failure in low-texture scenes such as track tunnels and single backgrounds. Furthermore, its high computational complexity makes it difficult to meet the real-time processing requirements of trains. The other type is based on motion sensor-assisted calibration, which uses devices such as accelerometers and gyroscopes to acquire motion parameters and apply them to image correction. However, traditional methods often employ simple linear displacement compensation models, which cannot cope with nonlinear image offsets under complex train motion conditions, especially for dynamic blurring and non-uniform afterimages caused by large acceleration changes. In addition, existing calibration methods generally do not distinguish between the target subject and the environment region in the image, applying a uniform calibration strategy to both, which easily leads to distortion of the target subject features. Simultaneously, there is a lack of dynamic optimization mechanisms for the calibration model. When the train's operating environment or motion state exceeds the preset range, the calibration accuracy will significantly decrease, making it difficult to adapt to the complex and variable nature of rail transit scenarios.

[0004] Therefore, in order to solve the problems existing in the prior art, the present invention proposes an image feature calibration method based on motion offset. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an image feature calibration method based on motion offset.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An image feature calibration method based on motion offset includes the following steps:

[0008] In the motion image acquisition step, the train uses encoders installed on the vehicle to record its speed information in real time, while cameras are positioned beside the track. Through a real-time vehicle-to-ground communication mechanism, the train transmits the speed information recorded by the encoders to a common unit near the camera. Upon receiving the speed information, the common unit analyzes the speed to determine the appropriate acquisition frequency, thereby acquiring motion images of the train in motion, while simultaneously reading ideal environment images from a historical database.

[0009] The initial image calibration step involves acquiring the acceleration at the time of train shooting, determining the basic displacement deviation of the image based on the acceleration, and performing preliminary image calibration on the motion image to obtain a reference processed image.

[0010] Blur feature calibration steps: Based on the target features in the historical database, identify the target subject and the environment region in the benchmark image, and extract the environment pixels within the environment region; combine the motion parameters of the train at the time of shooting and the ideal environment image to generate an environment point processing curve containing the expected features, extract the actual features of the environment pixels and perform feature-level fitting with the environment point processing curve, and select pixels with deviations exceeding a preset threshold as the pixel set to be calibrated.

[0011] The calibration image output step involves calibrating the pixel set to be calibrated using a fuzzy feature calibration strategy, and then fusing the calibrated pixel set with the reference processed image to obtain the processed standard motion image.

[0012] As a further improvement of the present invention, the initial image calibration step includes: acquiring real-time acceleration data at the time of shooting using an accelerometer mounted on the train, the real-time acceleration data including the magnitude and direction of acceleration; converting the real-time acceleration data into a basic displacement deviation of the image in the pixel dimension based on a preset acceleration-displacement conversion model, the basic displacement deviation including pixel offset values ​​in the horizontal and vertical directions; performing reverse compensation correction on the pixel coordinates of the moving image according to the basic displacement deviation; filling the pixel missing areas caused by displacement using neighbor pixel interpolation; and outputting the preliminarily calibrated image as a reference processing image.

[0013] As a further improvement of the present invention, the fuzzy feature calibration step includes: performing feature comparison on the benchmark processed image based on target features in a historical database, identifying and separating the target subject and the environment region in the image, extracting environment pixels from the environment region, wherein the environment pixels are pixels that are not the target subject in the benchmark processed image; combining the motion parameters at the time of train shooting and the ideal environment image, calculating and generating an environment point processing curve through an environment point processing strategy, wherein the environment point processing curve includes the expected spatial distribution features and grayscale change features of the environment pixels when there is no motion offset, wherein the motion parameters include the motion direction and the rate of change of acceleration; extracting the actual features of the environment pixels, wherein the actual features include spatial position offset, pixel grayscale attenuation rate and gradient direction; performing feature-level fitting between the actual features of the environment pixels and the expected features of the environment point processing curve, filtering out pixels whose actual features deviate from the expected features by a preset threshold, and outputting the set of pixels to be calibrated.

[0014] As a further improvement of the present invention, the ideal environment image is an environmental image acquired by the train under preset ideal operating conditions and after manual annotation and motion offset calibration. The image acquisition device on the train periodically acquires images when the train is under preset ideal operating conditions, and the acquisition frequency is not lower than a preset threshold. The target features are obtained by manually annotating the acquired ideal environment images, extracting the feature parameters of key objects, establishing a feature model and storing it in a historical database, and updating and optimizing the feature model to obtain a target feature database.

[0015] As a further improvement of the present invention, the environmental point processing curve is a multi-dimensional feature set describing the feature change law of environmental pixels under no motion offset state, including pixel spatial location distribution curve, gray value decay curve and gradient direction change curve. The environmental point processing curve extracts feature parameters of all environmental pixels in the environmental area based on ideal environmental images in the historical database, establishes a benchmark change curve of each feature parameter through statistical analysis, and obtains the motion parameters at the time of train shooting, including motion direction, instantaneous speed and acceleration change rate. The benchmark change curve is dynamically corrected based on a preset motion parameter feature influence model. The corrected curve is compared with the environmental image features without motion offset within a preset time in the historical database, the deviation is calculated and the curve parameters are optimized through an iterative algorithm to generate the environmental point processing curve under the current motion state.

[0016] As a further improvement of the present invention, the motion parameter feature influence model is obtained by training the historical data of train motion state and corresponding image feature changes through machine learning algorithm to obtain the mapping relationship between motion parameters and image feature changes, and the input parameters and output parameters of the model are determined. When the motion parameters at the time of train shooting are obtained, the motion parameter feature influence model is input to obtain the corresponding feature change correction coefficient, and the baseline change curve is corrected pixel by pixel according to the feature change correction coefficient to obtain the corrected environmental point processing curve.

[0017] As a further improvement of the present invention, the fuzzy feature calibration strategy includes: for each pixel in the pixel set to be calibrated, calculating a pixel position compensation value, a grayscale correction coefficient, and an edge sharpening parameter based on the deviation of the pixel from the environmental point processing curve, the train motion parameters, and the pixel features at the corresponding position in the ideal environmental image; performing coordinate offset compensation, grayscale value correction, and edge contour enhancement processing on the pixel according to the calibration parameters to generate calibrated pixels; setting a spatial distance weight factor to assign different fusion weights to the pixel to be calibrated and the surrounding uncalibrated pixels, with the weight ratio being higher the closer the distance; and fusing the calibrated pixels with the pixels in the corresponding region of the reference processed image using a weighted average algorithm, and performing smooth transition processing on the fusion boundary to generate a standard motion image.

[0018] As a further improvement of the present invention, an environmental point processing curve calibration step is also included. When the proportion of the pixel set to be calibrated exceeds a preset ratio threshold after feature fitting of several consecutive images, a preset number of ideal environmental images and corresponding motion parameters at the time of the image are selected from the historical database as calibration samples. The current environmental point processing curve is applied to the calibration samples, and the overall deviation value between the actual features of the environmental pixels in the samples and the expected features of the curve is calculated. Based on the overall deviation value, the feature parameters of the environmental point processing curve are corrected by the least squares method. The correction magnitude is positively correlated with the deviation value. The adjusted environmental point processing curve is applied to new verification samples. If the feature fitting deviation of the verification sample is lower than the preset qualified threshold, the curve is updated; otherwise, the curve calibration continues.

[0019] As a further improvement of the present invention, a calibration effect verification step is also included, which sets verification indicators, including pixel deviation rate, feature matching degree and image sharpness score, wherein the pixel deviation rate is the average pixel position deviation between the calibrated image and the ideal environment image, and the feature matching degree is the degree of overlap between the target subject features and the standard features in the historical database; the processed standard motion image is subjected to indicator detection, and the actual values ​​of each verification indicator are calculated; the actual values ​​are compared with the preset qualified thresholds, and if all meet the standards, the image is output; if not, the calibration step corresponding to the failed indicator is triggered to perform parameter backtracking adjustment, and the entire process from initial image calibration to image fusion is repeated until the verification indicators meet the standards.

[0020] The beneficial effects of this invention are:

[0021] (1) The basic displacement deviation is corrected based on the acceleration parameters through the initial image calibration step, and the target subject and the environment area are distinguished through the fuzzy feature calibration step. Feature-level fitting and screening are performed on the environmental pixels, and the calibration accuracy is layered. This effectively solves the problem of insufficient correction of nonlinear offset and dynamic afterimage by traditional methods, and significantly improves the detail restoration of motion images.

[0022] (2) Based on the ideal environment image and target feature model in the historical database, a dynamically adjusted environment point processing curve is generated by combining real-time motion parameters. The model is adaptively optimized through a special curve calibration step to ensure that the calibration effect remains stable when the train's motion state changes or the environment is different, thus overcoming the limitation of traditional methods that are highly dependent on fixed scenes.

[0023] (3) By comparing target features, the target subject and the environment area are separated, and the environmental pixels are calibrated in a targeted manner. This avoids interference with the features of key targets such as track markers and signal lights during the calibration process, ensuring the integrity and accuracy of the target subject features, and providing reliable image data support for subsequent track monitoring, safety warning and other applications. Attached Figure Description

[0024] Figure 1 This is a flowchart of an image feature calibration method based on motion offset according to the present invention;

[0025] Figure 2 This is a flowchart of the fuzzy feature calibration steps in an embodiment of the present invention;

[0026] Figure 3 This is a flowchart of the steps for constructing the environmental point processing curve according to an embodiment of the present invention;

[0027] Figure 4 This is a flowchart of the environmental point processing curve calibration steps according to an embodiment of the present invention;

[0028] Figure 5 This is a flowchart of the calibration effect verification steps in an embodiment of the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0030] This invention proposes an image feature calibration method based on motion offset, such as... Figures 1 to 5 As shown, it includes the following steps:

[0031] The motion image acquisition step involves acquiring motion images of the train during operation and reading ideal environment images from the historical database.

[0032] The train's onboard image acquisition device captures motion images in real time during operation. This device can be a high-definition industrial camera or a vision sensor, with a viewing angle covering the area in front of the train and its surroundings. Simultaneously, it retrieves ideal environment images matching the current route from a pre-built historical database. These ideal environment images are acquired and processed under preset ideal operating conditions, including uniform train speed, stationary state, and standard lighting conditions. These images undergo manual annotation and motion-off-free calibration, serving as a benchmark for image feature comparison.

[0033] The initial image calibration step involves acquiring the acceleration at the time of train shooting, determining the basic displacement deviation of the image based on the acceleration, and performing preliminary image calibration on the motion image to obtain a reference processed image.

[0034] Real-time acceleration data at the moment of image capture is acquired using an accelerometer mounted on the train. This data includes the specific magnitude and direction of the acceleration. Based on a pre-defined acceleration-displacement conversion model, the real-time acceleration data is converted into a basic displacement deviation in the pixel dimension of the image. This deviation includes pixel offset values ​​in both the horizontal and vertical directions. Inverse compensation correction is applied to the pixel coordinates of the moving image based on the basic displacement deviation. For pixel gaps caused by displacement, neighbor pixel interpolation is used to fill them. After coordinate correction and gap filling, a pre-calibrated image is output as the baseline image.

[0035] Blur feature calibration steps: Based on the target features in the historical database, identify the target subject and the environment region in the benchmark image, and extract the environment pixels within the environment region; combine the motion parameters of the train at the time of shooting and the ideal environment image to generate an environment point processing curve containing the expected features, extract the actual features of the environment pixels and perform feature-level fitting with the environment point processing curve, and select pixels with deviations exceeding a preset threshold as the pixel set to be calibrated.

[0036] First, feature comparison analysis is performed on the benchmark image based on target features from the historical database to identify and separate the target subject from the environmental region in the image. Target features are formed by manually annotating key objects in the ideal environmental image, extracting feature parameters, building a feature model, and optimizing it. Key objects include track markings, traffic lights, surrounding buildings, and other objects requiring special attention. Environmental pixels are extracted from the separated environmental region; these are pixels in the benchmark image that do not belong to the target subject. Combining the motion parameters of the train at the time of capture with the ideal environmental image, an environmental point processing curve is calculated using an environmental point processing strategy. Motion parameters include motion direction and acceleration rate of change. The environmental point processing curve includes the expected spatial distribution characteristics and grayscale change characteristics of environmental pixels in a state without motion offset. Actual features of the environmental pixels are extracted, including spatial position offset, pixel grayscale attenuation rate, and gradient direction. Feature-level fitting is performed between the actual features of the environmental pixels and the expected features of the environmental point processing curve. Pixels whose deviation between actual and expected features exceeds a preset threshold are selected and integrated to form a pixel set to be calibrated.

[0037] The calibration image output step involves calibrating the pixel set to be calibrated using a fuzzy feature calibration strategy, and then fusing the calibrated pixel set with the reference processed image to obtain the processed standard motion image.

[0038] A fuzzy feature calibration strategy is employed to calibrate pixels in the pixel set to be calibrated. For each pixel in the set, the pixel position compensation value, grayscale correction coefficient, and edge sharpening parameters are calculated by combining its deviation from the environmental point processing curve, train motion parameters, and pixel features at the corresponding position in the ideal environmental image. Based on these calibration parameters, coordinate offset compensation, grayscale value correction, and edge contour enhancement are performed on the pixels to generate calibrated pixels. A spatial distance weighting factor is set, and different fusion weights are assigned to the pixels to be calibrated and surrounding uncalibrated pixels according to this factor, with pixels closer to each other having a higher weight. A weighted average algorithm is used to fuse the calibrated pixels with the corresponding pixels in the reference image, while smoothing the fusion boundary to finally generate the processed standard motion image.

[0039] Specifically, such as Figures 1 to 5 As shown, the initial image calibration step includes: acquiring real-time acceleration data at the time of shooting using an accelerometer mounted on the train, the real-time acceleration data including the magnitude and direction of acceleration; converting the real-time acceleration data into a basic displacement deviation in the pixel dimension of the image based on a preset acceleration-displacement conversion model, the basic displacement deviation including pixel offset values ​​in the horizontal and vertical directions; performing reverse compensation correction on the pixel coordinates of the moving image according to the basic displacement deviation; filling the pixel missing areas caused by displacement using neighbor pixel interpolation; and outputting the preliminarily calibrated image as a reference processing image.

[0040] The train continuously collects real-time acceleration data at the moment of image capture using an accelerometer sensor. This sensor can be integrated into the train's onboard control system and can accurately capture the train's acceleration information during its movement. The acquired real-time acceleration data includes not only the specific magnitude of the acceleration but also the direction of the acceleration, thus comprehensively reflecting the changes in the train's motion state at the moment of image capture.

[0041] Real-time acceleration data is processed based on a pre-defined acceleration-displacement conversion model, converting physical acceleration parameters into basic displacement deviations in the image at the pixel level. This conversion model is established by analyzing a large amount of historical data on train motion states and corresponding image offsets, enabling the quantification of the mapping relationship between acceleration and pixel offset. The converted basic displacement deviations are specifically manifested as pixel offset values ​​in the horizontal and vertical directions, corresponding to the degree of image displacement in the lateral and longitudinal directions, respectively.

[0042] The pixel coordinates of the moving image are corrected using reverse compensation based on the calculated baseline displacement deviation. During the correction process, the coordinate position of each pixel is adjusted in the opposite direction to the displacement to counteract the initial offset caused by train motion. For pixel missing regions caused by pixel displacement during the correction process, a neighborhood pixel interpolation method is used for filling. Neighborhood pixel interpolation extracts feature information from the effective pixels surrounding the missing region and generates filling pixels according to spatial distribution rules to ensure the continuity of the image in the missing region. After coordinate correction and missing pixel filling, the image with preliminary calibration is output, which serves as the reference image for subsequent fine calibration.

[0043] Specifically, such as Figures 1 to 5As shown, the fuzzy feature calibration step includes: comparing the target features in the historical database with the benchmark image to identify and separate the target subject and the environment region in the image; extracting environment pixels from the environment region, where the environment pixels are pixels that are not the target subject in the benchmark image; calculating and generating an environment point processing curve by combining the motion parameters of the train at the time of shooting and the ideal environment image, where the environment point processing curve includes the expected spatial distribution features and grayscale change features of the environment pixels when there is no motion offset, where the motion parameters include the motion direction and the rate of change of acceleration; extracting the actual features of the environment pixels, where the actual features include spatial position offset, pixel grayscale attenuation rate, and gradient direction; performing feature-level fitting between the actual features of the environment pixels and the expected features of the environment point processing curve, filtering out pixels whose actual features deviate from the expected features by a preset threshold, and outputting the set of pixels to be calibrated.

[0044] Based on pre-constructed target features in a historical database, feature comparison analysis is performed on the benchmark processed image. Target features are formed through feature extraction and model training of key objects in the ideal environment image, covering feature parameters such as color, shape, and texture of the key objects. During feature comparison, feature information from the benchmark processed image is extracted and matched with target features to accurately identify the target subject requiring focus in the image, while separating other regions in the image into environment regions. Environment pixels are extracted from the separated environment regions; these pixels belong to parts of the benchmark processed image that do not belong to the target subject. The extraction process involves pixel-level region division of the image, and based on the boundary features between the target subject and the environment region, all pixels within the environment region are included in the extraction scope, forming an environment pixel set. Combining the motion parameters of the train at the time of shooting with the ideal environment image, an environment point processing curve is calculated and generated using an environment point processing strategy. The motion parameters specifically include the train's motion direction and rate of change of acceleration, reflecting the dynamic motion state of the train at the moment of shooting; the ideal environment image provides an environmental feature benchmark without motion offset. The environment point processing strategy integrates the influence of motion parameters on image features with the feature distribution of an ideal environment image to construct a multi-dimensional feature change model, thereby generating an environment point processing curve. This curve includes the expected spatial distribution features and expected grayscale change features of environment pixels in a state without motion offset. The expected spatial distribution features describe the ideal positional distribution of pixels in the image coordinate system, while the expected grayscale change features reflect the normal trend of pixel grayscale values ​​changing with spatial position. The actual features of the environment pixels are extracted, including spatial position offset, pixel grayscale attenuation rate, and gradient direction. The spatial position offset is obtained by calculating the difference between the actual coordinates and ideal coordinates of the environment pixel, reflecting the positional deviation caused by motion offset. The pixel grayscale attenuation rate is determined by analyzing the proportion of difference between the pixel grayscale value and the corresponding grayscale value in the ideal environment image, reflecting the degree of grayscale distortion caused by motion blur. The gradient direction is obtained by calculating the maximum direction of grayscale change in the pixel's neighborhood, used to characterize the actual direction of the pixel edge. The actual features of the environment pixels are then fitted with the expected features of the environment point processing curve at the feature level. The fitting process establishes feature mapping relationships, compares the degree of fit between actual and expected features one by one, and calculates the difference between the two. A preset threshold is set as a judgment criterion, and pixels whose difference between actual and expected features exceeds the threshold are selected. These pixels are determined to be regions significantly affected by motion offset, and are integrated to form a set of pixels to be calibrated and output.

[0045] Specifically, such as Figures 1 to 5As shown, the ideal environment image is an environmental image acquired by the train under preset ideal operating conditions and after manual annotation and motion offset calibration. The image acquisition device on the train periodically acquires images when the train is under preset ideal operating conditions, and the acquisition frequency is not lower than a preset threshold. The target features are obtained by manually annotating the acquired ideal environment images, extracting the feature parameters of key objects, establishing a feature model and storing it in a historical database, and updating and optimizing the feature model to obtain the target feature database.

[0046] The ideal environment image is captured when the train is under preset ideal operating conditions, and then manually annotated and calibrated to eliminate motion offset. The preset ideal operating conditions refer to scenarios where the train's operation is stable and external environmental interference is minimal, including situations where the train is traveling at a constant speed, stationary, or under standard lighting conditions. Under these conditions, image offset caused by train movement is minimal, and the captured images accurately reflect the original characteristics of the environment. Periodic image acquisition is performed using an image acquisition device mounted on the train to obtain the ideal environment image. The image acquisition device must have stable imaging performance to ensure the clarity and color consistency of the acquired images. The acquisition process is conducted continuously at fixed time intervals, with an acquisition frequency not lower than a preset threshold, to ensure that a sufficient number of sample images are obtained under different environmental conditions and road sections, providing sufficient data support for subsequent feature extraction and model construction. The target features are obtained through manual annotation of the acquired ideal environment images. During the annotation process, technicians accurately identify and select key objects in the images, including track markings, signal lights, roadside buildings, and other environmental elements requiring focused monitoring. Based on the annotation results, key object feature parameters are extracted. These parameters encompass information that characterizes the object's uniqueness, such as color distribution range, geometric features, and surface texture patterns. The extracted feature parameters are then integrated to build a feature model, which is stored in a historical database. To adapt to environmental changes and subtle alterations in object features, the feature model needs to be updated and optimized periodically. The update process involves introducing newly acquired ideal environment image data and adjusting the weights of the feature parameters in the model; the optimization process involves removing outlier samples and correcting feature biases to improve the model's accuracy in recognizing key objects. This continuously updated and optimized set of feature models collectively constitutes a target feature database, providing a stable and reliable benchmark for recognizing target subjects in images.

[0047] Specifically, such as Figures 1 to 5As shown, the environmental point processing curve is a multi-dimensional feature set describing the change law of environmental pixel features under no motion offset state, including pixel spatial location distribution curve, gray value decay curve and gradient direction change curve. The environmental point processing curve extracts feature parameters of all environmental pixels in the environmental area based on ideal environmental images in the historical database, establishes the benchmark change curve of each feature parameter through statistical analysis, and obtains the motion parameters at the time of train shooting, including motion direction, instantaneous speed and acceleration change rate. Based on the preset motion parameter feature influence model, the benchmark change curve is dynamically corrected. The corrected curve is compared with the environmental image features without motion offset within a preset time in the historical database, the deviation is calculated and the curve parameters are optimized through iterative algorithm to generate the environmental point processing curve under the current motion state.

[0048] The environmental point processing curve is a multi-dimensional feature set used to describe the variation law of environmental pixel features under no motion offset state. It includes pixel spatial location distribution curve, gray value decay curve, and gradient direction change curve. Among them, the pixel spatial location distribution curve is used to characterize the ideal position distribution law of environmental pixels in the image coordinate system, the gray value decay curve reflects the normal decay trend of environmental pixel gray value with spatial location change, and the gradient direction change curve describes the natural distribution characteristics of the gradient direction of environmental pixel edges. The three together constitute a complete characterization of the features of the offset-free environmental image. The generation process of the environmental point processing curve begins with ideal environmental images in the historical database. Feature parameters of all environmental pixels in the environmental area are extracted from the ideal environmental images. These parameters cover key information such as spatial coordinates, gray value, and gradient direction. These feature parameters are processed through statistical analysis to calculate the mean, distribution probability, and change trend of feature parameters under different spatial positions. Based on this, a baseline change curve for each feature parameter is established. The baseline change curve reflects the inherent change law of environmental pixel features under no motion interference. After obtaining the motion parameters at the time of train shooting, the baseline change curve is dynamically corrected. Motion parameters include the train's direction of motion, instantaneous speed, and rate of change of acceleration. These parameters accurately reflect the train's dynamic motion state at the moment of capture. Based on a pre-defined motion parameter feature influence model, motion parameters are transformed into feature correction coefficients. This model learns the correlation between historical motion data and image feature changes, establishing a quantitative mapping relationship between motion parameters and feature deviations. This allows for point-by-point adjustment of the feature parameters of the baseline curve based on real-time motion parameters, achieving dynamic curve correction. The corrected curve then undergoes further optimization. The corrected curve is compared with motion-off-free environmental image features within a preset time period in the historical database. These motion-off-free environmental images are reliable samples verified through prior calibration, reflecting the true state of recent environmental features. By calculating the deviation between the corrected curve and the sample features, an iterative algorithm is used to optimize and adjust the curve parameters. During the iteration process, the correction amplitude is dynamically adjusted according to the magnitude of the deviation, gradually reducing the difference between the curve and the sample features until the deviation falls within a preset range, ultimately generating an environmental point processing curve adapted to the current motion state.

[0049] Specifically, such as Figures 1 to 5 As shown, the motion parameter feature influence model is trained using a machine learning algorithm based on historical data of train motion state and corresponding image feature changes to obtain the mapping relationship between motion parameters and image feature changes. The model input parameters and output parameters are determined. When the motion parameters at the time of train shooting are obtained, the motion parameter feature influence model is input to obtain the corresponding feature change correction coefficient. The baseline change curve is then corrected pixel by pixel based on the feature change correction coefficient to obtain the corrected environmental point processing curve.

[0050] The construction of the motion parameter feature influence model begins with in-depth analysis of historical data. This historical data covers the motion state information of the train under different operating conditions, as well as the feature change data of the images collected under the corresponding conditions. Motion state information includes parameters such as acceleration, instantaneous speed, direction of motion, and rate of change of acceleration during train operation; image feature change data includes feature parameters such as pixel position offset, grayscale value change amplitude, and gradient direction deviation caused by motion shift. By training these historical data using machine learning algorithms, the algorithm can autonomously learn the intrinsic relationship between motion parameters and image feature changes, thereby establishing a quantitative mapping relationship between motion parameters and image feature changes. This mapping relationship is the core content of the motion parameter feature influence model. During model training, the input and output parameters of the model are clearly defined. Input parameters are selected as key parameters that comprehensively reflect the dynamic motion state of the train, including direction of motion, instantaneous speed, and rate of change of acceleration. These parameters directly affect the degree of offset and change pattern of image features. The output parameter is set as a feature change correction coefficient, which is used to quantify the degree of influence of motion parameters on image features, providing a precise adjustment basis for subsequent curve correction. Once the real-time motion parameters of the train at the moment of capture are obtained, these parameters are input into a pre-trained motion parameter feature influence model. The model calculates and processes the input motion parameters according to a pre-defined mapping relationship, outputting corresponding feature change correction coefficients. These correction coefficients are associated with pixel features at different spatial locations, specifically reflecting the differences in the impact of motion parameters on the features of each pixel in the image. Based on the feature change correction coefficients, the baseline change curve is corrected pixel by pixel. For each pixel, the expected value of the feature parameters is adjusted according to its corresponding correction coefficient, so that the baseline change curve can adapt to the image feature change patterns under the current train motion state, ultimately obtaining the corrected environmental point processing curve.

[0051] Specifically, such as Figures 1 to 5 As shown, the fuzzy feature calibration strategy includes: for each pixel in the pixel set to be calibrated, calculating the pixel position compensation value, grayscale correction coefficient, and edge sharpening parameters based on the deviation between the pixel and the environmental point processing curve, train motion parameters, and pixel features at the corresponding position in the ideal environmental image; performing coordinate offset compensation, grayscale value correction, and edge contour enhancement processing on the pixel according to the calibration parameters to generate calibrated pixels; setting a spatial distance weight factor to assign different fusion weights to the pixel to be calibrated and surrounding uncalibrated pixels, with a higher weight ratio for closer distances; fusing the calibrated pixels with the pixels in the corresponding region of the reference processed image using a weighted average algorithm, and performing smooth transition processing on the fusion boundary to generate a standard motion image.

[0052] The blur feature calibration strategy calculates calibration parameters for each pixel in the pixel set to be calibrated through multi-dimensional parameter analysis. The analysis process comprehensively considers the deviation between the pixel and the environmental point processing curve, the motion parameters at the time of train capture, and the pixel features at the corresponding position in the ideal environmental image. The deviation reflects the degree of deviation between the actual and expected features of the pixel, the motion parameters provide a reference for the dynamic motion state of the train, and the features of the ideal environmental image serve as a benchmark in a no-offset state. Based on this input information, three sets of key calibration parameters are calculated using a pre-set algorithm model: a pixel position compensation value to correct coordinate deviations caused by motion offset, a grayscale correction coefficient to adjust grayscale distortion caused by dynamic blur, and an edge sharpening parameter to enhance pixel edge features weakened by motion blur. Pixels are then processed specifically according to the calculated calibration parameters. Coordinate offset compensation adjusts offset pixels to their ideal spatial positions by superimposing position compensation values ​​onto their actual coordinates, restoring their correct distribution in the image. Grayscale correction adjusts pixel grayscale values ​​linearly or non-linearly based on grayscale correction coefficients, making the corrected grayscale values ​​closer to the true grayscale values ​​in the unoffset state. Edge contour enhancement strengthens the grayscale gradient of the pixel neighborhood based on edge sharpening parameters, improving the contrast of edge areas and making blurry contours clear. After these three processes, calibrated pixels are generated, whose features are closer to the ideal state without motion offset. To achieve natural fusion of calibrated pixels with the original image, a spatial distance weighting factor is set. This factor dynamically allocates fusion weights based on the spatial distance between pixels, assigning different weight values ​​to the pixel to be calibrated and its surrounding uncalibrated pixels. Among them, uncalibrated pixels closer to the pixel to be calibrated have a higher weight ratio to ensure a smooth transition of features in the fusion area; pixels farther away have a lower weight ratio to reduce interference with the core features of the calibrated area. A weighted average algorithm is used to fuse the calibrated pixels with the corresponding pixels in the benchmark image. During the fusion process, the final value of each pixel is calculated by weighting the calibrated pixel value with the surrounding uncalibrated pixel values ​​according to a weighted ratio. This preserves the accurate features after calibration and maintains the regional correlation of the original image. Simultaneously, the boundaries of the fusion region are smoothed through gradient adjustment techniques to weaken abrupt grayscale changes at the boundaries and eliminate stitching artifacts between different processed regions. After fusion and boundary processing, a standard motion image with coherent overall features and clear details is output.

[0053] Specifically, such as Figures 1 to 5As shown, the process also includes an environmental point processing curve calibration step. When the proportion of the pixel set to be calibrated exceeds a preset threshold after feature fitting of several consecutive images, a preset number of ideal environmental images and their corresponding motion parameters are selected from the historical database as calibration samples. The current environmental point processing curve is applied to the calibration samples, and the overall deviation between the actual features of the environmental pixels in the samples and the expected features of the curve is calculated. Based on the overall deviation, the feature parameters of the environmental point processing curve are corrected using the least squares method. The correction magnitude is positively correlated with the deviation value. The adjusted environmental point processing curve is applied to new verification samples. If the feature fitting deviation of the verification sample is lower than the preset qualified threshold, the curve is updated; otherwise, the curve calibration continues.

[0054] This step optimizes the environmental point processing curve when its accuracy deteriorates. Its activation mechanism is based on continuous monitoring of image feature fitting results: when the proportion of the pixel set to be calibrated in the total pixels of the environmental area exceeds a preset threshold after feature fitting of several consecutively acquired images, the system determines that the adaptability of the current environmental point processing curve has decreased and triggers the curve calibration process. The first step in curve calibration is selecting calibration samples. A preset number of ideal environmental images are retrieved from the historical database. These images have undergone rigorous verification to ensure no motion offset and reliable features. Simultaneously, train motion parameters corresponding to the acquisition time of these images are extracted, including direction of motion, instantaneous speed, and rate of change of acceleration, forming a calibration sample set. The calibration samples must cover typical scenes under different road sections and lighting conditions to ensure comprehensive calibration. The currently used environmental point processing curve is applied to the selected calibration samples, and the overall deviation between the actual features of the environmental pixels in the samples and the expected features of the curve is calculated through feature comparison. The overall deviation value is calculated by comprehensively considering the feature deviations of all environmental pixels in the samples, comprehensively reflecting the degree of deviation between the current curve and the actual environmental features, providing a quantitative basis for subsequent corrections. Based on the calculated overall deviation value, the feature parameters of the environmental point processing curve are corrected using the least squares method. The least squares method constructs an optimization model that minimizes the sum of squared deviations to find the parameter adjustment values ​​that best approximate the sample features. During the correction process, the adjustment magnitude of the feature parameters is positively correlated with the overall deviation value; that is, the larger the deviation, the greater the parameter adjustment magnitude, ensuring that the curve can specifically compensate for the deviation. After correction, the effectiveness of the adjusted environmental point processing curve needs to be verified. The adjusted curve is applied to a new verification sample, which is an ideal environmental image and its corresponding motion parameters that have not been calibrated. The fitting deviation between the environmental pixel features and the expected curve features in the verification sample is calculated. If the deviation is lower than a preset acceptable threshold, it indicates that the curve accuracy has been restored, and the environmental point processing curve in the system is updated. If the deviation does not reach the acceptable threshold, the deviation calculation step is returned, and the curve feature parameters are corrected again until the verification deviation meets the requirements.

[0055] Specifically, such as Figures 1 to 5 As shown, the process also includes a calibration effect verification step, which sets verification indicators, including pixel deviation rate, feature matching degree, and image sharpness score. The pixel deviation rate is the average pixel position deviation between the calibrated image and the ideal environment image, and the feature matching degree is the overlap between the target subject features and the standard features in the historical database. The processed standard moving image is tested for the indicators, and the actual values ​​of each verification indicator are calculated. The actual values ​​are compared with the preset qualified thresholds. If all indicators meet the standards, the image is output. If the indicators do not meet the standards, the calibration step corresponding to the failed indicator is triggered to backtrack and adjust the parameters. The entire process from initial image calibration to image fusion is repeated until the verification indicators meet the standards.

[0056] This step is used to quantitatively evaluate and optimize the quality of the processed standard moving images. First, multi-dimensional verification indicators are set, including pixel deviation rate, feature matching degree, and image sharpness score. The pixel deviation rate is calculated by the average positional deviation of all pixels in the calibrated image compared to their corresponding pixels in the ideal environment image, reflecting the overall spatial positioning calibration accuracy of the image. The feature matching degree is determined by comparing the overlap between the target subject features and standard features in the historical database, reflecting the accuracy of key object feature restoration. The image sharpness score is generated by analyzing the image edge gradient and the richness of detail texture, evaluating the overall visual quality of the image. Indicator detection is performed on the processed standard moving images. The actual coordinates of each pixel in the image are extracted using image analysis algorithms and compared with the corresponding pixel coordinates in the ideal environment image to calculate the pixel deviation rate. Feature comparison technology is used to match the feature parameters of the target subject with standard features in the historical database, and the proportion of overlapping features is counted to obtain the feature matching degree. A sharpness evaluation model is used to quantitatively analyze the edge sharpness and texture details of the image, generating an image sharpness score. Through these detection processes, the actual values ​​of each verification indicator are obtained. The actual values ​​of each indicator are then compared with preset pass thresholds. The acceptable threshold is set based on the accuracy requirements of the rail transit image application scenario to ensure that the image quality meets the requirements of subsequent monitoring, recognition, and other applications. If the actual values ​​of all indicators reach or exceed the acceptable threshold, it means that the image calibration effect meets expectations, and the standard motion image is directly output. If there are indicators that do not meet the standard, the corresponding calibration steps are triggered according to the type of indicator that failed to pass, and parameters are adjusted retrospectively. For example, if the pixel deviation rate does not meet the standard, the displacement compensation parameters in the initial image calibration are adjusted retrospectively; if the feature matching degree is insufficient, the feature fitting threshold in the blur feature calibration is optimized. After the adjustment is completed, the entire process from initial image calibration to image fusion is re-executed until all verification indicators reach the acceptable threshold.

[0057] The foregoing has illustrated and described the basic features, principles, and advantages of the present invention. It should be noted that the present invention is not limited to the above embodiments, but only to some embodiments. Any improvements and additions made without departing from the spirit and scope of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. An image feature calibration method based on motion offset, characterized in that, Includes the following steps: The motion image acquisition step involves acquiring motion images of the train during operation and reading ideal environment images from the historical database. The initial image calibration step involves acquiring the acceleration at the time of train shooting, determining the basic displacement deviation of the image based on the acceleration, and performing preliminary image calibration on the motion image to obtain a reference processed image. Blur feature calibration steps: Based on the target features in the historical database, identify the target subject and the environment region in the benchmark image, and extract the environment pixels within the environment region; combine the motion parameters of the train at the time of shooting and the ideal environment image to generate an environment point processing curve containing the expected features, extract the actual features of the environment pixels and perform feature-level fitting with the environment point processing curve, and select pixels with deviations exceeding a preset threshold as the pixel set to be calibrated. The calibration image output step involves calibrating the pixel set to be calibrated using a fuzzy feature calibration strategy, and then fusing the calibrated pixel set with the reference processed image to obtain the processed standard motion image. The environmental point processing curve is a multi-dimensional feature set describing the change law of environmental pixel features under no motion offset state, including pixel spatial location distribution curve, gray value decay curve and gradient direction change curve. The environmental point processing curve extracts feature parameters of all environmental pixels in the environmental area based on ideal environmental images in the historical database, establishes the baseline change curve of each feature parameter through statistical analysis, and obtains the motion parameters at the time of train shooting, including motion direction, instantaneous speed and acceleration change rate. Based on the preset motion parameter feature influence model, the baseline change curve is dynamically corrected. The corrected curve is compared with the environmental image features without motion offset within a preset time in the historical database, the deviation is calculated and the curve parameters are optimized through iterative algorithm to generate the environmental point processing curve under the current motion state. The fuzzy feature calibration strategy includes, for each pixel in the pixel set to be calibrated, calculating the pixel position compensation value, grayscale correction coefficient, and edge sharpening parameter based on the deviation between the pixel and the environmental point processing curve, the train motion parameters, and the pixel features at the corresponding position in the ideal environmental image; and performing coordinate offset compensation, grayscale value correction, and edge contour enhancement processing on the pixel according to the calibration parameters to generate calibrated pixels. A spatial distance weighting factor is set, and different fusion weights are assigned to the pixels to be calibrated and the surrounding uncalibrated pixels. The closer the distance, the higher the weight ratio. The calibrated pixels are fused with the corresponding pixels in the benchmark image through a weighted average algorithm, and the fusion boundary is smoothed to generate a standard motion image.

2. The image feature calibration method based on motion offset according to claim 1, characterized in that, The initial image calibration step includes: acquiring real-time acceleration data at the moment of shooting using an accelerometer mounted on the train, the real-time acceleration data including the magnitude and direction of acceleration; converting the real-time acceleration data into a basic displacement deviation in the pixel dimension of the image based on a preset acceleration-displacement conversion model, the basic displacement deviation including pixel offset values ​​in the horizontal and vertical directions; performing reverse compensation correction on the pixel coordinates of the moving image according to the basic displacement deviation; filling the pixel missing areas caused by displacement using neighbor pixel interpolation; and outputting the preliminarily calibrated image as a reference image.

3. The image feature calibration method based on motion offset according to claim 1, characterized in that, The fuzzy feature calibration step includes: comparing the target features in the historical database with the benchmark image to identify and separate the target subject and the environment region in the image; extracting environment pixels from the environment region, where the environment pixels are pixels that are not the target subject in the benchmark image; calculating and generating an environment point processing curve by combining the motion parameters of the train at the time of shooting and the ideal environment image, where the environment point processing curve includes the expected spatial distribution features and grayscale change features of the environment pixels when there is no motion offset, and the motion parameters include the motion direction and the rate of change of acceleration; extracting the actual features of the environment pixels, where the actual features include spatial position offset, pixel grayscale attenuation rate, and gradient direction; performing feature-level fitting between the actual features of the environment pixels and the expected features of the environment point processing curve, and filtering out pixels whose actual features deviate from the expected features by a preset threshold, outputting the pixel set to be calibrated.

4. The image feature calibration method based on motion offset according to claim 1, characterized in that, In the historical database, the ideal environment image is an environmental image collected by the train under preset ideal operating conditions and after manual annotation and motion offset calibration. The image acquisition device on the train periodically acquires images when the train is under preset ideal operating conditions, and the acquisition frequency is not lower than a preset threshold. The target features are obtained by manually annotating the acquired ideal environment images, extracting the feature parameters of key objects, establishing a feature model and storing it in the historical database, and updating and optimizing the feature model to obtain the target feature database.

5. The image feature calibration method based on motion offset according to claim 1, characterized in that, The motion parameter feature influence model is trained using a machine learning algorithm based on historical data of train motion state and corresponding image feature changes to obtain the mapping relationship between motion parameters and image feature changes. The model's input and output parameters are determined. When the motion parameters at the time of train shooting are obtained, the motion parameter feature influence model is input to obtain the corresponding feature change correction coefficient. The baseline change curve is then corrected pixel by pixel based on the feature change correction coefficient to obtain the corrected environmental point processing curve.

6. The image feature calibration method based on motion offset according to claim 1, characterized in that, It also includes an environmental point processing curve calibration step. When the proportion of the pixel set to be calibrated exceeds a preset threshold after feature fitting of several consecutive images, a preset number of ideal environmental images and their corresponding motion parameters are selected from the historical database as calibration samples. The current environmental point processing curve is applied to the calibration samples, and the overall deviation between the actual features of the environmental pixels in the samples and the expected features of the curve is calculated. Based on the overall deviation, the feature parameters of the environmental point processing curve are corrected using the least squares method. The correction magnitude is positively correlated with the deviation value. The adjusted environmental point processing curve is applied to new verification samples. If the feature fitting deviation of the verification sample is lower than the preset qualified threshold, the curve is updated; otherwise, the curve calibration continues.

7. The image feature calibration method based on motion offset according to claim 1, characterized in that, It also includes a calibration effect verification step, setting verification indicators, including pixel deviation rate, feature matching degree, and image sharpness score. The pixel deviation rate is the average pixel position deviation between the calibrated image and the ideal environment image, and the feature matching degree is the overlap between the target subject features and the standard features in the historical database. The processed standard moving image is tested for indicators, and the actual values ​​of each verification indicator are calculated. The actual values ​​are compared with the preset qualified thresholds. If all indicators meet the standards, the image is output. If the indicators do not meet the standards, the calibration step corresponding to the failed indicator is triggered to backtrack and adjust the parameters. The entire process from initial image calibration to image fusion is repeated until the verification indicators meet the standards.

Images