Image measurement vehicle roll angle method for structured pavement conditions

Abstract

The present application proposes a method for measuring the roll angle of an image measuring vehicle under structured road conditions, aiming to compensate for the cumulative error and shock effect of low-cost IMU measurement. The method first measures the geometric distance between laser beams through an improved image processing method, and then realizes high-precision roll angle measurement based on a regression function. In order to further improve the accuracy of measurement, the present application inputs the collected laser images into an improved MobileNetV4 model, optimizes the feature extraction and dynamic prediction ability through technologies such as multi-scale convolution, adaptive convolution kernel, dynamic feature selection and weighting, and thus realizes the prediction of the roll angle of the vehicle. By combining the two methods, using a weighted average fusion mechanism and dynamically adjusting the weight, the accuracy and stability of the roll angle measurement are further improved. The method can maintain high accuracy under structured road conditions, and compared with the low-cost IMU sensor measurement method, it has good stability and real-time performance. This research provides a new means for monitoring the dynamic behavior of vehicles, and has important significance for the development of vehicle safety and driving assistance systems.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle posture monitoring, and in particular to a method for measuring vehicle roll angle based on an image of a laser beam geometric distance under structured road conditions. Background Art

[0002] Vehicle rollover is an extremely dangerous traffic accident, especially for heavy vehicles such as trucks, which are prone to rollover when turning due to their high center of gravity and heavy load. Rollover accidents not only cause serious property damage, but may also endanger the lives of drivers and passengers. Studies have shown that the occurrence of vehicle rollover accidents is closely related to multiple factors, including the vehicle's driving state, road conditions, and vehicle posture. In anti-rollover technology, vehicle stability and posture control are key, and rollover prediction and warning have become the core of achieving effective prevention and control. Modern active safety control systems such as active suspension and differential braking avoid rollover accidents by adjusting the vehicle state in real time.

[0003] Real-time monitoring of the vehicle's roll angle plays a crucial role in these control systems. As a key parameter for assessing vehicle dynamic behavior and stability, roll angle directly indicates whether the vehicle is approaching the tipping point, enabling the system to promptly adjust the vehicle's driving state to prevent rollover. Accurate roll angle measurement is particularly important for improving vehicle stability and ensuring driving safety under extreme operating conditions, such as high-speed cornering and sharp obstacle avoidance.

[0004] Currently, vehicle roll angle measurement primarily relies on various sensors, including accelerometers and gyroscopes. While these sensors can provide measurement results, they are often limited by cost, installation complexity, and environmental adaptability. First, the integration and calibration of accelerometers and gyroscopes is relatively complex, which not only increases installation and maintenance costs but also places high demands on the installation location and vehicle structure. Furthermore, over extended periods of operation, these sensors are susceptible to environmental factors such as temperature and humidity, causing drift and reduced measurement accuracy. This is especially true when the vehicle is in long-term operation, undergoes frequent acceleration and deceleration, or encounters complex road conditions. This further exacerbates sensor error accumulation, making it difficult to ensure stable and accurate roll angle measurement. These issues limit the application of traditional sensors in roll angle measurement, making it difficult to maintain accurate and reliable measurements in complex driving environments.

[0005] Therefore, how to overcome the limitations of existing technologies and achieve accurate, stable and real-time roll angle measurement has become a key challenge to improve the performance of anti-rollover control systems. Summary of the Invention

[0006] To address the limitations of existing technologies, this paper proposes a method for measuring vehicle roll angle using images based on laser beam geometric distance, suitable for structured road conditions. Compared to existing technologies, this method demonstrates significant advantages in long-term measurement accuracy and environmental adaptability. It provides more stable and accurate measurement results for applications requiring high-precision, real-time roll angle information, and has broad application prospects.

[0007] The present invention adopts the following technical solutions:

[0008] A method for measuring vehicle roll angle based on an image of a laser beam geometric distance under structured road conditions includes the following:

[0009] Step S1: Mount two laser emitters on the outside of the vehicle's suspension base along the transverse axis, so that they form parallel linear light spots on the ground. Mount an imaging device (e.g., a camera) near one of the laser emitters. Adjust the position of the laser emitter so that the beam is within the imaging device's field of view and parallel to the vehicle's longitudinal axis. The imaging device continuously captures images of the laser beam on the ground to detect the vehicle's roll angle in real time.

[0010] Step S2: Mount the laser transmitter and imaging equipment on a platform, simulating the vehicle suspension configuration. During the experiment, the platform is tilted to simulate vehicle roll, and the tilt angle is recorded using a tilt sensor. The collected laser images are calibrated. Repeat the experiment under common structured road conditions to build a more comprehensive dataset, facilitating inclination measurements suitable for structured road conditions.

[0011] Step S3: Based on the data set established in step S2, the collected images are processed based on the Canny edge detection and Hough transform algorithms to calculate the geometric distance between the two laser beams. A regression model of the tilt angle-laser beam distance is then constructed and fitted using the least squares method to accurately predict the vehicle roll angle under different road conditions.

[0012] Step S4: To improve the accuracy and environmental adaptability of tilt angle detection, the present invention predicts the vehicle's roll angle using a modified MobileNetV4 model. This model utilizes adaptive convolution kernels and a dynamic feature selection mechanism to enhance its ability to process complex ground reflection information. It also supports quantization and pruning techniques, making it lightweight and easy to deploy in real time on vehicle platforms.

[0013] Step S5. Finally, the present invention utilizes a fusion mechanism based on weighted average to combine the S3 image processing with the prediction results of the S4 deep learning model, and dynamically adjusts the weights through adaptive correction to improve the accuracy and stability of vehicle roll angle measurement.

[0014] Furthermore, the specific steps of step S1 are as follows:

[0015] Step S11, laser light source arrangement: two laser emitters are installed horizontally side by side on the outside of the bottom of the vehicle body along the transverse axis of the vehicle, ensuring that the laser beams they emit can form clear, parallel linear light spots on the ground.

[0016] Step S12: Imaging Unit Setup: Select an imaging device (e.g., a camera) and install it near one of the laser emitters, ensuring that the device is between 30 and 50 centimeters above the ground. The imaging device should support RGB 24-bit color images and have a resolution of at least 480×320 to ensure image quality.

[0017] Step S13, light spot positioning: adjust the positions of the two laser emitters so that the linear light spots they form on the ground are within the field of view of the imaging device, and ensure that each light line is parallel to the longitudinal axis of the vehicle.

[0018] Step S14, image capture: In order to detect the roll angle of the vehicle in real time, the imaging device continuously captures the image of the laser spot on the ground at a speed of about 30 frames per second.

[0019] Furthermore, the specific steps of step S2 are as follows:

[0020] Step S21: After installing the laser transmitter and imaging equipment, an experimental setup is constructed to collect laser road image data under common structured road conditions. To safely simulate vehicle roll while moving, the present invention uses a platform in place of the vehicle's suspension. The laser transmitter and imaging equipment are mounted on the platform, ensuring that the mounting positions are consistent with those on an actual vehicle. Furthermore, a high-precision tilt sensor is installed beneath the platform for dataset calibration.

[0021] Step S22: During the experiment, the platform was tilted to simulate vehicle roll, with the roll angle controlled between -15 and 15 degrees. By adjusting the platform's tilt angle, nearly 80 laser road surface images were captured under the same road surface conditions. The corresponding angles of these images were evenly distributed between -15 and 15 degrees. The roll angles recorded in real time by the tilt sensor were used to calibrate the collected image data to ensure the accuracy of the dataset.

[0022] Step S23: After completing the dataset for a single road surface type, the experimental scope was further expanded to cover common structured road conditions: asphalt, mud, cement, and gravel. For each road surface type, corresponding laser road surface images were systematically collected according to established methods. This method creates a comprehensive dataset that represents the various road conditions a vehicle might encounter during actual driving, providing a solid data foundation for accurate vehicle roll angle measurement.

[0023] Furthermore, the specific steps of step S3 are as follows:

[0024] First, the images collected by the imaging device are converted into grayscale images to reduce data dimensionality and improve processing efficiency. Next, to remove noise and detail interference from the images, the present invention uses adaptive Gaussian blurring technology and dynamically adjusts smoothing parameters to further improve image quality, providing high-quality input data for subsequent feature extraction.

[0025] Next, the present invention uses the Canny edge detection algorithm to extract edge information in the image. The specific method is as follows: first, the gradient of the pre-processed image is calculated, and the area with the most obvious brightness change in the image is identified by calculating the brightness change in the horizontal and vertical directions, that is, the area most likely to be the edge of the laser beam, and a non-maximum suppression strategy is adopted to screen out the strongest local edges to avoid repeated identification and blurred edge responses. The edges are classified through a dual-threshold strategy: that is, high-intensity edges are immediately identified, and low-intensity edges are determined to be valid through their connection relationship with high-intensity edges. This dual-threshold mechanism effectively suppresses the noise contours in the background, ensuring that only clear edges related to the target features are retained. Through the edge detection algorithm, the output edge image only contains edges with structural significance, laying the foundation for the subsequent laser beam feature extraction.

[0026] Edge detection can remove most of the noise, but it still includes some contours in the ground background. In order to accurately extract the two laser beams, the present invention first analyzes the edge detection results and further eliminates isolated points or small edge areas. This step calculates the area or perimeter of each edge area and excludes those areas that are smaller than expected or do not conform to the characteristics of the laser beam based on the set threshold, thereby reducing background interference and highlighting the laser beam features. The Hough transform is then used to extract the laser beam features in the image. The Hough transform maps each edge point in the image to the parameter space and identifies the point set that conforms to the straight line equation. The specific steps are as follows:

[0027] Polar coordinate space mapping: For each edge point, according to its image coordinates (x, y), it is converted into line parameters (ρ, θ) in polar coordinates, where

[0028] ρ=x·cos(θ)+y·sin(θ)

[0029] In this process, each edge point will form a curve in the ρ-θ parameter space.

[0030] When multiple edge points are collinear, the curves they form in ρ-θ space intersect at the same point. This intersection represents a line in the image. By accumulating the number of occurrences of these intersections, we can find the points in parameter space that appear the most frequently. These points correspond to significant lines in the image.

[0031] To ensure detection accuracy, the present invention incorporates a dynamic threshold adjustment mechanism into the Hough transform, retaining only those points whose cumulative counts in the parameter space exceed a certain threshold. This threshold is dynamically adjusted based on the specific application environment and laser beam intensity, ensuring that only significant straight lines are detected while ignoring irrelevant background contours.

[0032] After the line detection is completed, multiple lines may be detected in the image. In order to accurately extract the two laser beams, the present invention uses the following strategies for screening and optimization:

[0033] (1) Redundant line detection: First, the distance between each pair of detected lines is calculated, usually based on their ρ parameter (i.e., the distance between the line and the image origin). For lines with a distance that is too small, it is judged to be a redundant detection result of the same line; then, the θ angle difference between each two lines is calculated. If the angle difference between the two lines is less than a set threshold and their distance difference is also within a certain range, they are considered to be the same line and merged. This step generates a more stable line parameter by calculating the mean or median.

[0034] (2) Laser Beam Screening: To ensure that only two laser beams are extracted, the present invention further screens the lines based on their geometric position, length, and relationship to the laser emitter's emission direction. Based on a pre-set geometric model, the two lines that meet the expected position and direction of the laser beam can be quickly identified, while other background lines that do not meet the conditions are excluded.

[0035] After completing the image processing, the equations for the centerlines of the two laser beams are obtained. The horizontal axis is then set to the actual vehicle roll angle, which is recorded by the tilt sensor during image acquisition. The vertical axis is set to the pixel position of the centerline of the two laser beams. This is the horizontal position of the centerline in the image calculated after the image processing step. A curve is then fitted to these data points using the least squares method. The least squares method determines the best-fit curve by minimizing the sum of squared errors, thereby establishing a regression model that accurately represents the relationship between the vehicle roll angle and the centerline pixel position.

[0036] Further, the specific steps of the step S4 are as follows:

[0037] After the construction of the basic regression model is completed, in order to further improve the accuracy and environmental adaptability of the vehicle roll angle measurement system, the present application predicts the roll angle of the vehicle through an improved MobileNetV4 model. MobileNetV4 adopts an inverted bottleneck block (IRB) structure, effectively reducing the parameter quantity and calculation quantity of the model through depth separable convolution, while maintaining good performance. In addition, MobileNetV4 further enhances the model's ability to capture multi-scale features through the use of hierarchical interconnection and feature fusion strategies.

[0038] In order to better adapt to the specific needs of vehicle roll angle measurement, especially to provide stable and reliable measurement results under structured road conditions, the present application improves the MobileNetV4. First, the adaptive convolution kernel mechanism is introduced, which enables the model to dynamically adjust the parameters of the convolution kernel according to the shape and intensity of the laser beam in the input image, thereby more accurately extracting features related to the roll angle. This optimization significantly improves the model's ability to process complex ground reflection information, enhancing the accuracy of the prediction and adaptability under different road conditions.

[0039] Secondly, considering the resource limitations that the model may face in actual deployment, the present application fully considers the compatibility of quantization and pruning techniques when designing the network structure. By selecting appropriate activation functions and data representation methods for quantization, and constructing a network structure that is easy to prune, the optimized MobileNetV4 model can reduce the volume and computational resource consumption of the model while maintaining high-precision prediction, enabling real-time processing on the vehicle-mounted computing platform. This feature enables the present application to perform real-time data fusion with the vehicle's existing IMU system, providing more accurate and stable roll angle measurement results.

[0040] In addition, the present application also introduces dynamic feature selection and attention mechanisms. In the feature fusion stage, the model can learn the importance of different features to the prediction of the roll angle, and select and weight the features according to these importance. This mechanism enables the model to pay more attention to features that have a significant impact on the prediction result, while ignoring irrelevant or noisy features, thereby further improving the accuracy of the prediction. This dynamic feature selection capability enables the present application to perform well in long-term measurement, effectively compensating for the cumulative errors that may occur in the IMU over a long period of use.

[0041] The improved MobileNetV4 network produces a highly accurate and environmentally adaptable vehicle roll angle measurement model. It accurately analyzes various structured road scenarios, improving prediction accuracy through adaptive convolution kernels and dynamic feature selection. Its lightweight design also facilitates deployment on in-vehicle computing platforms. This deep learning approach complements traditional IMU measurements, providing more stable and accurate roll angle information, particularly in situations with frequent acceleration and deceleration or long driving periods, providing reliable data support for vehicle stability control and safety systems.

[0042] Furthermore, the specific steps of step S5 are as follows:

[0043] After completing the design of the MobileNetV4 model, the present invention integrates the model with image processing methods to further improve the accuracy and environmental adaptability of vehicle roll angle measurement. The specific integration steps are as follows:

[0044] First, the present invention introduces a fusion mechanism based on weighted averaging to combine the advantages of image processing methods and deep learning models. This mechanism dynamically adjusts the weights of image processing results and deep learning predictions based on current road conditions, lighting conditions, and other external environmental factors. For scenes with good lighting and smooth roads, the image processing method is given a higher weight. However, in scenes with complex lighting and rough roads, the deep learning model is given a higher weight to account for complex backgrounds and noise interference.

[0045] Subsequently, at each time step, the weighted average method is used to fuse the image processing results and the deep learning prediction results to generate a final roll angle prediction value.

[0046] Assume that the prediction result of the image processing method is θ img , the prediction result of the deep learning model is θ dl , then the final roll angle prediction value can be calculated by the following formula:

[0047] θ final =w img (t)·θ img +w dl (t)·θ dl

[0048] To achieve adaptive dynamic adjustment, the present invention also introduces an adaptive correction mechanism to further optimize the accuracy of the fusion results. This mechanism analyzes the fusion results over multiple time steps and dynamically adjusts the weights. In this strategy, the system adjusts the weights of the image processing method's predictions and those of the deep learning model in real time based on changes in the external environment, such as light intensity, road surface roughness, and vehicle speed.

[0049] According to the roll angle prediction formula above, its weight value is updated by the following formula:

[0050]

[0051]

[0052] Here, α is the adjustment coefficient, which controls the system's sensitivity to errors. A larger α value will make the system more sensitive to errors, resulting in more drastic weight changes; while a smaller α value will make the weight adjustment smoother.

[0053] The normalization factor Z(t) is used to ensure that the sum of the two weights is 1, and is calculated as follows:

[0054] Z(t)=exp(-α·|θ img (t)-θ final (t)|)+exp(-α·|θ dl (t)-θ final (t)|)

[0055] In this way, the system can automatically adjust the weights according to the prediction accuracy of different methods at each time step. For example, when the prediction error of the image processing method is small, w img (t+1) will increase, which means that the system will rely more on the results of image processing. On the contrary, if the prediction error of the deep learning model is small, w dl (t+1) will increase, making the system more dependent on the predictions of the deep learning model.

[0056] The beneficial effect of this invention lies in that this integrated approach not only achieves highly accurate geometric analysis but also improves adaptability to common structured road conditions. This technology provides real-time prediction of vehicle roll angles and effectively compensates for the accumulated errors of IMUs over long-term measurements. Compared to existing technologies, this invention demonstrates significant advantages in long-term measurement accuracy and environmental adaptability, providing an efficient and reliable complementary solution for vehicle stability control and road safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the vehicle roll angle measurement system of the present invention.

[0058] Figure 2 This is a schematic diagram of the installation of the inclination measurement of the present invention.

[0059] Figure 3 Schematic diagram of the image processing technology of the present invention. DETAILED DESCRIPTION

[0060] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0061] Figure 1 This is a flow chart of the vehicle roll angle measurement system of the present invention. The present invention will be further explained below with reference to the accompanying drawings.

[0062] Step 1: Install two laser emitters on the outside of the vehicle's suspension bottom along the transverse axis, so that they form parallel linear light spots on the ground. Mount an imaging device (such as a camera) near one of the laser emitters. Adjust the position of the laser emitter so that the beam is within the imaging device's field of view and parallel to the vehicle's longitudinal axis. The imaging device continuously captures images of the laser beam on the ground to detect the vehicle's roll angle in real time.

[0063] Step 2: Mount the laser transmitter and imaging equipment on a platform, simulating the vehicle suspension configuration. During the experiment, the platform was tilted to simulate vehicle roll, and the tilt angle was recorded using a tilt sensor. The collected laser images were calibrated. The experiment was repeated under common structured road conditions to build a more comprehensive dataset, enabling inclination measurements suitable for structured road conditions.

[0064] Step 3. Based on the dataset established in step 2, the collected images are processed using the Canny edge detection and Hough transform algorithms to calculate the geometric distance between the two laser beams. A regression model of the inclination angle-laser beam distance is then constructed and fitted using the least squares method to accurately predict the vehicle roll angle under different road conditions.

[0065] Step 4: To improve the accuracy and environmental adaptability of tilt detection, this paper predicts the vehicle's roll angle using a modified MobileNetV4 deep learning model. This model uses adaptive convolution kernels and a dynamic feature selection mechanism to enhance its ability to process complex ground reflection information. It also supports quantization and pruning techniques, making it lightweight and easy to deploy in real time on vehicle platforms.

[0066] Step 5. Finally, the present invention utilizes a fusion mechanism based on weighted average to combine the image processing in step 3 with the prediction results of the deep learning model in step 4, and dynamically adjusts the weights through adaptive correction to improve the accuracy and stability of vehicle roll angle measurement.

[0067] The schematic diagram of the inclination measurement installation in step 1 is as follows Figure 2 As shown, the specific implementation is:

[0068] First, the laser light source is arranged: two laser emitters are mounted horizontally side by side along the vehicle's transverse axis, outside the vehicle's suspension base. The laser beams they emit form a clear, parallel linear spot on the ground. Next, the imaging unit is set up: an imaging device (such as a camera) is selected and mounted close to one of the laser emitters, ensuring it is between 30 and 50 centimeters above the ground. The imaging device should capture RGB 24-bit color images and have a resolution of at least 480×320 to ensure image quality. Next, beam positioning is performed: the two laser emitters are positioned so that the linear spot they form on the ground is within the imaging device's field of view, ensuring that each beam is parallel to the vehicle's longitudinal axis. Finally, image capture is performed: to detect the vehicle's roll angle in real time, the imaging device continuously captures images of the laser beams on the ground at a rate of approximately 30 frames per second.

[0069] The specific implementation of step 2 is:

[0070] After installing the laser transmitter and imaging equipment, an experimental setup was constructed to collect laser road image data under common structured road conditions. To safely simulate vehicle roll while driving, a platform was used in place of the vehicle's suspension. The laser transmitter and imaging equipment were mounted on the platform, ensuring that their installation positions were consistent with those on an actual vehicle. Furthermore, a high-precision tilt sensor was installed beneath the platform for dataset calibration.

[0071] During the experiment, the platform was tilted to simulate vehicle roll, with the roll angle controlled between -15 and 15 degrees. By adjusting the platform's tilt angle, nearly 80 laser road surface images were captured under the same road conditions, with the corresponding angles uniformly distributed between -15 and 15 degrees. The roll angles recorded in real time by the tilt sensor were used to calibrate the collected image data to ensure the accuracy of the dataset.

[0072] After completing the dataset for a single road surface type, the experiment was further expanded to cover common structured road conditions: asphalt, mud, cement, and gravel. For each road surface type, corresponding laser road surface images were systematically collected according to established methods. This approach enabled the construction of a comprehensive dataset that represents the various road conditions that vehicles may encounter during real-world driving, providing a solid data foundation for accurate vehicle roll angle measurement.

[0073] The process diagram of the image processing technology in step three is as follows Figure 3 As shown, the specific implementation is:

[0074] First, the images collected by the imaging device are converted into grayscale images to reduce data dimensionality and improve processing efficiency. To remove noise and detail interference in the image, the present invention uses adaptive Gaussian blur technology and dynamically adjusts the smoothing parameters to further improve image quality, providing high-quality input data for subsequent feature extraction.

[0075] Next, the present invention uses the Canny edge detection algorithm to extract edge information in the image. The specific method is as follows: first, the gradient of the pre-processed image is calculated, and the area with the most obvious brightness change in the image is identified by calculating the brightness change in the horizontal and vertical directions, that is, the area most likely to be the edge of the laser beam, and a non-maximum suppression strategy is adopted to screen out the strongest local edges to avoid repeated identification and blurred edge responses. The edges are classified through a dual-threshold strategy: that is, high-intensity edges are immediately identified, and low-intensity edges are determined to be valid through their connection relationship with high-intensity edges. This dual-threshold mechanism effectively suppresses the noise contours in the background, ensuring that only clear edges related to the target features are retained. Through the edge detection algorithm, the output edge image only contains edges with structural significance, laying the foundation for the subsequent laser beam feature extraction.

[0076] Edge detection can remove most of the noise, but it still includes some contours in the ground background. In order to accurately extract the two laser beams, the present invention first analyzes the edge detection results and further eliminates isolated points or small edge areas. This step calculates the area or perimeter of each edge area and excludes those areas that are smaller than expected or do not conform to the characteristics of the laser beam based on the set threshold, thereby reducing background interference and highlighting the laser beam features. The Hough transform is then used to extract the laser beam features in the image. The Hough transform maps each edge point in the image to the parameter space and identifies the point set that conforms to the straight line equation. The specific steps are as follows:

[0077] Polar coordinate space mapping: For each edge point, according to its image coordinates (x, y), it is converted into line parameters (ρ, θ) in polar coordinates, where

[0078] ρ=x·cos(θ)+y·sin(θ)

[0079] In this process, each edge point will form a curve in the ρ-θ parameter space.

[0080] When multiple edge points are collinear, the curves they form in ρ-θ space intersect at the same point. This intersection represents a line in the image. By accumulating the number of occurrences of these intersections, we can find the points in parameter space that appear the most frequently. These points correspond to significant lines in the image.

[0081] To ensure detection accuracy, the present invention incorporates a dynamic threshold adjustment mechanism into the Hough transform, retaining only those points whose cumulative counts in the parameter space exceed a certain threshold. This threshold is dynamically adjusted based on the specific application environment and laser beam intensity, ensuring that only significant straight lines are detected while ignoring irrelevant background contours.

[0082] After the line detection is completed, multiple lines may be detected in the image. In order to accurately extract the two laser beams, the present invention uses the following strategies for screening and optimization:

[0083] (1) Redundant line detection: First, the distance between each pair of detected lines is calculated, and the difference is calculated based on their ρ parameter (i.e., the distance between the line and the image origin). For lines with too small a distance, they are judged to be redundant detection results of the same line; then, the θ angle difference between each two lines is calculated. If the angle difference between the two lines is less than a certain set threshold and their distance difference is also within a certain range, they are considered to be the same line and merged. This step generates a more stable line parameter by calculating the mean or median.

[0084] (2) Line Feature Screening: To ensure that only two laser beams are extracted, the present invention further screens the lines based on their geometric position, line length, and their relationship to the laser emitter's emission direction. Based on a pre-set geometric model, the two lines that match the expected position and direction of the laser beam can be quickly identified, while other background lines that do not meet the conditions are excluded.

[0085] After completing the image processing, the equations for the centerlines of the two laser beams are obtained. The horizontal axis is then set to the actual vehicle roll angle, which is recorded by the tilt sensor during image acquisition. The vertical axis is set to the pixel position of the centerline of the two laser beams. This is the horizontal position of the centerline in the image calculated after the image processing step. A curve is then fitted to these data points using the least squares method. The least squares method determines the best-fit curve by minimizing the sum of squared errors, thereby establishing a regression model that accurately represents the relationship between the vehicle roll angle and the centerline pixel position.

[0086] The specific implementation of step 4 is:

[0087] After completing the construction of the basic regression model, in order to further improve the accuracy and environmental adaptability of the vehicle roll angle measurement system, this paper predicts the vehicle roll angle using an improved MobileNetV4 model. MobileNetV4 adopts an inverted bottleneck block (IRB) structure and effectively reduces the model's parameter count and computational complexity through depthwise separable convolution, while maintaining good performance. In addition, MobileNetV4 further enhances the model's ability to capture multi-scale features by adopting a hierarchical interconnection approach and feature fusion strategy.

[0088] To better meet the specific needs of vehicle roll angle measurement, particularly providing stable and reliable results under varying road conditions, this paper improves upon MobileNetV4. First, an adaptive convolution kernel mechanism is introduced, enabling the model to dynamically adjust the kernel parameters based on the shape and intensity of the laser beam in the input image, thereby more accurately extracting roll angle-related features. This optimization significantly improves the model's ability to handle complex ground reflection information, enhancing prediction accuracy and adaptability to varying road conditions.

[0089] Secondly, considering the resource constraints that the model may face in actual deployment, the present invention fully considers the compatibility of quantization and pruning technologies when designing the network structure. By selecting activation functions and data representations suitable for quantization, and constructing a network structure that is easy to prune, the optimized MobileNetV4 model can maintain high-precision predictions while reducing model size and computing resource consumption, enabling real-time processing on the vehicle's computing platform. This feature enables real-time data fusion with the vehicle's existing IMU system, providing more accurate and stable roll angle measurements.

[0090] Furthermore, this invention incorporates a dynamic feature selection and attention mechanism. During the feature fusion stage, the model learns the importance of different features to roll angle prediction and selects and weights them accordingly. This mechanism enables the model to focus on features that significantly influence the prediction while ignoring irrelevant or noisy features, further improving prediction accuracy. This dynamic feature selection capability enables this invention to excel in long-term measurements, effectively compensating for the accumulated errors that may occur with long-term IMU use.

[0091] The improved MobileNetV4 network produces a highly accurate and environmentally adaptable vehicle roll angle measurement model. It accurately analyzes common structured road scenarios, improving prediction accuracy through adaptive convolution kernels and dynamic feature selection. Its lightweight design also facilitates deployment on in-vehicle computing platforms. This deep learning approach complements traditional IMU measurements, providing more stable and accurate roll angle information, particularly in situations with frequent acceleration and deceleration or long driving periods, providing reliable data support for vehicle stability control and safety systems.

[0092] The specific implementation of step five is:

[0093] After completing the design of the MobileNetV4 model, the present invention integrates the model with the image processing method to further improve the accuracy and environmental adaptability of vehicle roll angle measurement. The specific integration steps are as follows:

[0094] First, the present invention introduces a fusion mechanism based on weighted averaging to combine the advantages of image processing methods and deep learning models. This mechanism dynamically adjusts the weights of image processing results and deep learning predictions based on current road conditions, lighting conditions, and other external environmental factors. For scenes with good lighting and smooth roads, the image processing method is given a higher weight. However, in scenes with complex lighting and rough roads, the deep learning model is given a higher weight to account for complex backgrounds and noise interference.

[0095] Subsequently, at each time step, the weighted average method is used to fuse the image processing results and the deep learning prediction results to generate a final roll angle prediction value.

[0096] Assume that the prediction result of the image processing method is θ img , the prediction result of the deep learning model is θ dl , then the final roll angle prediction value can be calculated by the following formula:

[0097] θ final =w img (t)·θ img +w dl (t)·θ dl

[0098] To achieve adaptive dynamic adjustment, the present invention also introduces an adaptive correction mechanism to further optimize the accuracy of the fusion results. This mechanism analyzes the fusion results over multiple time steps and dynamically adjusts the weights. In this strategy, the system adjusts the weights of the image processing method's predictions and those of the deep learning model in real time based on changes in the external environment, such as light intensity, road surface roughness, and vehicle speed.

[0099] According to the roll angle prediction formula above, its weight value is updated by the following formula:

[0100]

[0101]

[0102] Here, α is the adjustment coefficient, which controls the system's sensitivity to errors. A larger α value will make the system more sensitive to errors, resulting in more drastic weight changes; while a smaller α value will make the weight adjustment smoother.

[0103] The normalization factor Z(t) is used to ensure that the sum of the two weights is 1, and is calculated as follows:

[0104] Z(t)=exp(-α·|θ img (t)-θ final (t)|)+exp(-α·|θ dl (t)-θ final (t)|)

[0105] In this way, the system can automatically adjust the weights according to the prediction accuracy of different methods at each time step. For example, when the prediction error of the image processing method is small, w img (t+1) will increase, which means that the system will rely more on the results of image processing. On the contrary, if the prediction error of the deep learning model is small, w dl (t+1) will increase, making the system more dependent on the predictions of the deep learning model.

[0106] Ultimately, this adaptive weight adjustment mechanism ensures the system automatically optimizes the fusion results under varying environmental and road conditions, improving the overall accuracy and environmental adaptability of the vehicle's roll angle prediction. By integrating real-time data with the IMU, the proposed method effectively compensates for the IMU's accumulated errors, providing more reliable and accurate roll angle information for the vehicle stability control system.

[0107] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for measuring vehicle roll angle using images on structured road conditions, characterized by: The following steps are involved: Step S1: Two laser emitters are mounted on the outside of the bottom of the vehicle suspension along the transverse axis so that they form parallel linear light spots on the ground; an imaging device is mounted near one of the laser emitters; the position of the laser emitter is adjusted so that the light beam is within the field of view of the imaging device and parallel to the longitudinal axis of the vehicle; the imaging device continuously captures images of the laser beam on the ground to detect the vehicle's roll angle in real time; Step S2: Mounting the laser transmitter and imaging device on a platform to simulate the installation configuration of a vehicle suspension; During the experiment, the platform is tilted to simulate vehicle roll, and the tilt angle is recorded using a tilt sensor, and the collected laser images are calibrated; Repeating the experiment under common structured road conditions to establish a more comprehensive data set to facilitate the completion of tilt angle measurements suitable for structured road conditions; Step S3: Based on the data set established in step S2, the collected images are processed using Canny edge detection and Hough transform algorithms to calculate the geometric distance between the two laser beams, and a relationship model between the tilt angle and the laser beam distance is constructed. The least squares method is used for fitting to accurately predict the vehicle roll angle under structured road conditions. Step S4: To improve the accuracy and environmental adaptability of tilt angle detection, the vehicle's roll angle is predicted using an improved MobileNetV4 model. The improved MobileNetV4 model uses adaptive convolution kernels and a dynamic feature selection mechanism to enhance its ability to process complex ground reflection information. It also supports quantization and pruning techniques, making it lightweight and easy to deploy in real time on vehicle platforms. Step S5: Using a fusion mechanism based on weighted average, the image processing in step S3 is combined with the prediction results of the deep learning model in step S4, and the weights are dynamically adjusted through adaptive correction to improve the accuracy and stability of vehicle roll angle measurement.

2. The method for measuring vehicle roll angle using images on structured road conditions according to claim 1, characterized in that: The specific steps of step S1 are as follows: Step S11, laser light source arrangement: two laser emitters are installed horizontally side by side on the outside of the vehicle suspension bottom along the transverse axis of the vehicle to ensure that a clear, parallel linear light spot is formed on the ground; Step S12, Imaging Unit Setup: Select an imaging device and install it near one of the laser emitters, ensuring that the vertical height of the device is between 30 and 50 centimeters from the ground; the imaging device supports capturing RGB 24-bit color images and has a resolution of at least 480×320 to ensure image quality; Step S13, light spot positioning: adjust the positions of the two laser emitters so that the linear light spots formed by them on the ground are within the field of view of the imaging device, and ensure that each light beam is parallel to the longitudinal axis of the vehicle; Step S14, image capture: In order to detect the roll angle of the vehicle in real time, the imaging device continuously captures the image of the laser beam on the ground at a speed of about 30 frames per second.

3. The method for measuring vehicle roll angle using images on structured road conditions according to claim 1, characterized in that: The specific steps of step S2 are as follows: Step S21: After the laser transmitter and imaging device are installed, an experimental setup is constructed to collect laser road image data under common structured road conditions. To simulate a rolling vehicle while ensuring safety, a platform is used instead of the vehicle's suspension. The laser transmitter and imaging device are installed on the platform, ensuring that the installation positions are consistent with those on the actual vehicle. In addition, a high-precision tilt sensor is installed under the platform for calibration of the data set; Step S22: During the experiment, the platform is tilted to simulate vehicle roll, and the roll angle is controlled between -15 degrees and 15 degrees. By adjusting the platform's tilt angle, 80 laser road surface images are captured under the same road surface conditions, and the angles corresponding to these images are evenly distributed between -15 degrees and 15 degrees. The roll angle recorded in real time by the tilt sensor is used to calibrate the collected image data to ensure the accuracy of the data set. Step S23: After completing the establishment of a dataset for a single road surface type, the experimental scope is further expanded to cover other structured road conditions: asphalt, mud, cement, and gravel roads. For each road surface type, corresponding laser road surface images are systematically collected according to the established method. Through this method, a comprehensive dataset is constructed that can represent various road conditions that a vehicle may encounter during actual driving, thereby providing a solid data foundation for accurate measurement of vehicle roll angle.

4. The method for measuring vehicle roll angle using images on structured road conditions according to claim 1, characterized in that: The specific steps of step S3 are as follows: First, the images collected by the imaging device are converted into grayscale images to reduce data dimensionality and improve processing efficiency. To remove noise interference in the image, adaptive Gaussian blur technology is used to dynamically adjust the smoothing parameters to further improve image quality and provide high-quality input data for subsequent feature extraction. Next, the Canny edge detection algorithm is used to extract edge information from the image. The specific method is as follows: First, the gradient of the preprocessed image is calculated. By calculating the brightness changes in the horizontal and vertical directions, the area with the most obvious brightness change in the image is identified, that is, the area most likely to be the edge of the laser beam. The non-maximum suppression strategy is used to screen out the strongest local edges to avoid repeated identification and blurred edge responses. The edges are classified using a dual-threshold strategy: high-intensity edges are immediately identified, and low-intensity edges are determined to be valid based on their connection relationship with high-intensity edges. This dual-threshold mechanism effectively suppresses noise contours in the background, ensuring that only clear edges related to the target features are retained. Through the edge detection algorithm, the output edge image only contains edges with structural significance, laying the foundation for the subsequent laser beam feature extraction. Edge detection removes most of the noise, but still includes some contours from the ground background. To accurately extract the two laser beams, the edge detection results are first analyzed to further eliminate isolated points or small edge areas. This step reduces background interference and highlights the characteristics of the laser beam by calculating the area or perimeter of each edge area and excluding areas smaller than expected or not in line with the laser beam characteristics based on a set threshold. The Hough transform is then used to extract the laser beam features in the image. The Hough transform maps each edge point in the image to a parameter space and identifies the point set that conforms to the straight line equation. The specific steps are as follows: Polar coordinate space mapping: For each edge point, according to its image coordinates (x, y), it is converted into line parameters (ρ, θ) in polar coordinates, where ρ=x·cos(θ)++y·sin(θ) In this process, each edge point will form a curve in the ρ-θ parameter space; When multiple edge points are collinear, the curves they form in the ρ-θ space intersect at the same point; this intersection represents a straight line in the image; by accumulating the number of occurrences of these intersections, we find the points that appear the most times in the parameter space, which correspond to the significant lines in the image; To ensure detection accuracy, a dynamic threshold adjustment mechanism is introduced into the Hough transform, retaining only points whose cumulative number of occurrences in the parameter space exceeds a certain threshold. This threshold is dynamically adjusted based on the specific application environment and the intensity of the laser beam, ensuring that only significant straight lines are detected while ignoring irrelevant background contours. After line detection is completed, multiple lines may be detected in the image. To accurately extract the two laser beams, the following strategies are used for screening and optimization: (1) Redundant line detection: First, the distance between each pair of detected lines is calculated, and the difference is calculated based on their ρ parameters. For lines with a distance that is too small, it is judged as a redundant detection result of the same line. Then, the angle difference θ between each pair of lines is calculated. If the angle difference between the two lines is less than a set threshold and the distance difference between them is also within a certain range, they are considered to be the same line and merged. This step generates a more stable line parameter by calculating the mean or median. (2) Laser beam screening: To ensure that only two laser beams are extracted, the lines are further screened based on their geometric position, line length, and relationship with the emission direction of the laser emitter. Based on the preset geometric model, the two lines that meet the expected position and direction of the laser beam are quickly identified, and other background lines that do not meet the conditions are excluded. After completing the image processing step, the centerline equations of the two laser beams are obtained; then, the horizontal axis is set to the actual roll angle of the vehicle. These angle data are synchronously recorded by the tilt sensor during image acquisition; the vertical axis is set to the pixel position of the centerline of the two laser beams, which is the horizontal position of the centerline in the image calculated after the image processing step; then, the least squares method is used to fit the curve to these data points; the least squares method determines the best fit curve by minimizing the sum of the squares of the errors, thereby establishing a regression model that can accurately represent the relationship between the vehicle's roll angle and the centerline pixel position.

5. The method for measuring vehicle roll angle using images on structured road conditions according to claim 2, characterized in that: The specific steps of step S4 are as follows: To further improve the accuracy and environmental adaptability of the vehicle roll angle measurement system, an improved MobileNetV4 model was used to predict the vehicle roll angle. MobileNetV4 uses an inverted bottleneck block (IRB) structure and depthwise separable convolution to effectively reduce the model's parameter count and computational complexity while maintaining good performance. Furthermore, MobileNetV4 further enhances the model's ability to capture multi-scale features through hierarchical interconnection and feature fusion strategies. To better meet the needs of vehicle roll angle measurement, especially providing stable and reliable measurement results on structured roads, improvements have been made to MobileNetV4. First, an adaptive convolution kernel mechanism has been introduced, enabling the model to dynamically adjust the convolution kernel parameters based on the shape and intensity of the laser beam in the input image, thereby more accurately extracting features related to the roll angle. This optimization significantly improves the model's ability to process complex ground reflection information, enhancing prediction accuracy and adaptability to structured road conditions. Secondly, considering the resource constraints faced by the model in actual deployment, the compatibility of quantization and pruning technologies was fully considered when designing the network structure. By selecting activation functions and data representation methods suitable for quantization, and constructing a network structure that is easy to prune, the optimized MobileNetV4 model can reduce the model size and computing resource consumption while maintaining high-precision predictions, thereby enabling real-time processing on the on-board computing platform. This feature enables real-time data fusion with the vehicle's existing IMU system, providing more accurate and stable roll angle measurement results. In addition, dynamic feature selection and attention mechanism are introduced; During the feature fusion phase, the model learns the importance of different features to roll angle prediction and selects and weights features based on their importance. This mechanism enables the model to focus more on features that have a significant impact on the prediction results, while ignoring those that are irrelevant or noisy, further improving prediction accuracy. This dynamic feature selection capability enables the present invention to perform well in long-term measurements, effectively compensating for the cumulative errors that may occur in the long-term use of the IMU. Through the improved MobileNetV4 network, a high-precision and environmentally adaptable vehicle roll angle measurement model is obtained; it can accurately analyze structured road scenes, improve prediction accuracy through adaptive convolution kernels and dynamic feature selection, and based on the lightweight model design, it can facilitate the model's deployment on the vehicle computing platform. This deep learning method complements traditional IMU measurements, especially in situations where the vehicle frequently accelerates and decelerates or travels for long periods of time, and can provide more stable and accurate roll angle information, providing reliable data support for vehicle stability control and safety systems.

6. The method for measuring vehicle roll angle using images applicable to structured road conditions according to claim 3, characterized in that: The specific steps of step S5 are as follows: After optimizing the MobileNetV4 model, we integrated it with image processing methods to further improve the accuracy and environmental adaptability of vehicle roll angle measurement. The specific integration steps are as follows: First, a fusion mechanism based on weighted averaging is introduced to combine the advantages of image processing methods and deep learning models. In this mechanism, the weights of image processing results and deep learning prediction results are dynamically adjusted according to the current road conditions, lighting conditions, and other external environmental factors. For scenes with good lighting conditions and smooth roads, image processing methods have a higher weight. However, in scenes with complex lighting and rough roads, the weight of deep learning models increases accordingly to cope with complex backgrounds and noise interference. Then, at each time step, the weighted average method is used to fuse the image processing results and the deep learning prediction results to generate a final roll angle prediction value; Assume that the prediction result of the image processing method is θ img , the prediction result of the deep learning model is θ dl , then the final roll angle prediction value θ final It can be calculated by the following formula: i final =w img (t)·θ img +w dl (t)·θ dl To achieve adaptive dynamic adjustment, an adaptive correction mechanism is introduced to further optimize the accuracy of the fusion results. This mechanism analyzes the fusion results of multiple consecutive time steps and dynamically adjusts the weight values. In this strategy, the system adjusts the weights of the prediction results of the image processing method and the deep learning model in real time based on changes in the external environment, including light intensity, road surface smoothness, and vehicle speed. According to the roll angle prediction formula, its weight value W img and w dl They are updated by the following formulas respectively: Among them, α is the adjustment coefficient, which controls the sensitivity of the system to errors. A larger α value will make the system more sensitive to errors, resulting in more drastic changes in weights; while a smaller α value will make the weight adjustment smoother. The normalization factor Z(t) is used to ensure that the sum of the two weights is 1, and is calculated as follows: Z(t)=exp(-α·|θ img (t)-θ final (t)|)+exp(-α·|θ dl (t)-θ final (t)|) In this way, the system can automatically adjust the weights according to the prediction accuracy of different methods at each time step; when the prediction error of the image processing method is small, w img (t+1) will increase, which means that the system will rely more on the results of image processing; on the contrary, if the prediction error of the deep learning model is small, w dl (t+1) will increase, making the system more dependent on the predictions of the deep learning model; Ultimately, this adaptive weight adjustment mechanism ensures that the system can automatically optimize the fusion results under different environmental and road conditions, improving the overall accuracy and environmental adaptability of the vehicle roll angle prediction; through real-time fusion with IMU data, it can effectively compensate for the IMU's accumulated errors and provide more reliable and accurate roll angle information for the vehicle stability control system.

Images