Vehicle-mounted road surface damage rapid detection method based on machine vision

Abstract

The invention discloses a vehicle-mounted road surface damage rapid detection method based on machine vision, and relates to the technical field of road surface detection, and the method comprises the following steps: S1, collecting vibration data; s2, processing the vibration data; s3, extracting an abnormal event; s4, classifying the abnormal events; s5, acquiring a pavement image; and S6, identifying the pavement damage type. The vehicle-mounted vibration detection system is arranged to measure the vehicle vibration, the acquisition speed and the measurement precision of the sensor can be improved, signal denoising is performed according to the characteristics of the vehicle vibration signals acquired by the sensor, recognition and distinguishing are performed according to waveform data generated by different vibrations, abnormal events can be rapidly and accurately extracted, and the accuracy of the abnormal events is improved. According to different energy characteristics corresponding to different abnormal events, energy of discrete signals is obtained through characteristic extraction, characteristics of non-pavement damage such as deceleration strips can be eliminated, and the condition of pavement damage can be accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of road surface detection, and in particular to a method for rapid detection of vehicle-mounted road surface damage based on machine vision. Background Art

[0002] The consequences of road damage include affecting driving quality and safety, reducing the service capacity and flatness of the road surface, causing uneven driving, increasing vehicle fuel consumption and tire wear, and also affecting driving comfort and safety. In addition, road damage can also cause rainwater penetration, affecting the structural strength of the road surface. Therefore, it is necessary to detect road damage. In the Chinese patent application number 202011146400.5, "a road damage detection device, a road damage detection method and a storage medium are disclosed. The road damage detection device is configured to calculate the maximum change rate for each road section, and the maximum change rate is the maximum value of the change in wheel speed per unit time in each vehicle. The device is also configured to periodically select the maximum value from the maximum change rates of each vehicle within a specified period of time for each road section, set the selected maximum value as the section maximum change rate, and determine whether the road damage has occurred by comparing the determination target value with the positive threshold value. The determination target value is obtained by subtracting a comparison value based on at least one of the section maximum change rates set before the last time from the latest set section maximum change rate."

[0003] The above-mentioned documents can achieve the purpose of road damage detection. However, during the entire road damage detection process, since road damage includes different types of damage such as potholes and bumps, as well as interference factors such as speed bumps and pits, it is difficult to distinguish whether it is a normal road phenomenon by simply calculating the maximum change rate of each road section. Therefore, it is difficult to extract accurate road damage data, which affects the final judgment. In addition, the characteristics of road damage are not accurately extracted, and the location, degree and type of road damage cannot be obtained, which is not convenient for rapid road maintenance and affects normal use. Summary of the invention

[0004] The purpose of the present invention is to provide a method for rapid detection of vehicle-mounted road damage based on machine vision to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for rapid detection of vehicle-mounted road surface damage based on machine vision, comprising the following steps:

[0006] S1. Collecting vibration data: Install the vehicle-mounted vibration detection system on the vehicle. When the vehicle encounters a pothole road during driving, the vehicle body will vibrate. The vehicle vibration is automatically measured during driving. The vehicle-mounted vibration detection system automatically sends the vibration data and geographic location information to the terminal software through the wireless transmission network, and the terminal software processes the data and determines the road damage;

[0007] S2. Processing vibration data: Input a section of vibration data, first perform maximum point filtering to remove abnormal point data that affects the average value judgment, then perform trend item filtering and empirical mode decomposition (EMD) denoising to filter out trend changes caused by small displacements of the vehicle-mounted vibration detection system and filter out high-frequency small signal background noise;

[0008] S3, extracting abnormal events: the input vibration data has different vibration types, including normal events, abnormal events and interference events. The interference events for abnormal event identification are filtered out by the average value method, and the abnormal events are extracted from the denoised data by the mean square value method;

[0009] S4. Classification of abnormal events: Different categories of abnormal events have different signal energies. The features of abnormal events are extracted by extracting IMF energy features, which are used as inputs of the SVM classifier to classify abnormal events, thereby eliminating vibrations caused by non-road damage in abnormal events;

[0010] S5. Collecting road images: A camera is installed on the vehicle. When an abnormal event caused by road damage is identified, the camera is triggered to collect road images. The collected images are grayed and enhanced. After that, the input images are converted into feature maps using a convolutional neural network. After the feature maps are extracted and sampled by multiple convolutional layers and pooling layers, the fully connected layer flattens the extracted final feature maps for target classification operations.

[0011] S6. Identify the type of road damage: Use the improved Yolov5 algorithm to detect the input feature map. During the detection, a large number of target detection frames are generated. When there is a road damage target, the image of the damaged part of the road is cut out, the degree of road damage is analyzed, the type of road damage is identified, and the analysis results are transmitted to the central detection part.

[0012] Preferably, in step S1, the vehicle-mounted vibration detection system is as follows: a vibration node is provided in the vehicle body, the vibration node includes an inductive sensor and a mass block connected to a spring, the vibration node is fixed on the vehicle body, when the vehicle body is used as the vibrating object to be measured, the space inside the vehicle is regarded as a reference system for vibration measurement, the vehicle body vibrates, the inductive sensor and the vehicle body are relatively stationary, and the spring drives the mass block to generate vibration different from the object to be measured, relative vibration is generated between the probe of the inductive sensor and the mass block, and the vibration condition of the object to be measured is reflected by measuring the amplitude of the relative vibration;

[0013] The entire measurement process includes a vehicle-mounted part and a central detection part. The vehicle-mounted part is used to detect, collect, and transmit vibration data and geographic location information, while the central detection part is used to receive data from the vehicle-mounted vibration detection system and identify the type of road damage.

[0014] Preferably, step S2 comprises the following steps:

[0015] S201, the maximum value point filtering process is to segment a section of noisy data, find the maximum value point of each section of data, and calculate the average value of each section of data, and replace the found maximum value point with the average value;

[0016] S202. Use the least squares method to filter out the trend term. Empirical mode decomposition (EMD) denoising uses the EMD signal decomposition algorithm to perform time-frequency decomposition based on the local characteristics of the signal itself to obtain results of decomposition at different frequencies. A low-pass filter is constructed based on the results to filter the background noise of the remaining high-frequency small signals in the signal.

[0017] Preferably, in step S202, the EMD signal decomposition algorithm is specifically:

[0018]

[0019] Among them, X(t) represents the original signal, IMF i (t) represents the IMF component. Each of the n IMF components represents a signal component of a frequency in the original signal. n represents the total number of IMF components. n (t) represents the residual component, t represents time, i represents the IMF component number, i∈[1,n];

[0020] The low-pass filter construction process is as follows: decompose the noisy signal to obtain IMF and residual components, calculate the signal energy of each IMF component and the residual component, then find the minimum value of the IMF component signal energy, add one to the minimum value as the dividing point, select the dividing point and the IMF components after it to reconstruct the denoised signal.

[0021] Preferably, in step S3, the normal event refers to the waveform generated by the vehicle running over a smooth and healthy road surface, the interference event refers to the waveform generated when the vehicle does not vibrate but the vibration node is detected to be shaking, including vehicle speed changes, braking and turning, and the abnormal event refers to the waveform generated by the vehicle running over a damaged road surface and the obvious vibration such as running over artificial obstacles such as speed bumps.

[0022] Preferably, in step S3, when an interference event occurs, the mass block in the vibration node generates an extremely low-frequency change in the equilibrium position, and the data generated by the vibration node deviating from the equilibrium position is obtained by average value calculation, and the data whose absolute value of the average value deviation exceeds a set threshold is determined as an interference event, and the data whose absolute value does not exceed the set threshold is determined as an abnormal event and a normal event;

[0023] The process of selecting the average value threshold is as follows: select three types of signal waveform data, namely normal events, interference events and abnormal events, for average value analysis, and calculate the absolute value of the average value deviation of each event. The absolute value of the average value deviation refers to the difference between the true value of the signal and the average value, and obtain the result of statistical analysis of the selected data. The threshold is set according to the result of statistical analysis.

[0024] Calculate the mean square value of each data segment, and determine the data whose mean square value deviation exceeds the set threshold as an abnormal event, otherwise it is a normal event;

[0025] The mean square value threshold selection process is as follows: two types of signal waveform data, normal events and abnormal events, are selected for mean square value analysis, and the mean square value deviation of each event is calculated to obtain the results of statistical analysis of the selected data, and the threshold is set according to the results of the statistical analysis.

[0026] Preferably, in step S4, the abnormal event data is decomposed by the EMD signal decomposition algorithm in step S2 to obtain n IMF components and residual components, and the ratio of the energy of each IMF component to the signal energy is used as the feature vector. The specific steps are as follows:

[0027] S401, calculate the energy E of each time domain IMF component i ;

[0028]

[0029] Among them, E i represents the energy of each time domain IMF component, c i (t) represents the energy of the IMF component of a single time domain, N represents the total time domain, t represents the time, i represents the time sequence number, n represents the last time, and I represents the time sequence number;

[0030] S402, constructing a vector T with the energy of each time-domain IMF component, and obtaining an IMF normalized energy feature T' through normalization processing;

[0031] T=[E1 E2···E n ];

[0032]

[0033] Where T represents the energy vector of each time domain IMF component, T' represents the IMF normalized energy feature, E represents the signal energy, and E i Represents the total energy of each time domain IMF component, i represents the numerical sequence, i∈[1,n], and n represents the total time domain.

[0034] Preferably, in step S6, the target detection frame includes the position of the target detection frame and the confidence information of the target detection frame, and the target detection frame with low confidence is eliminated by using a non-maximum suppression method, and then the acquired road surface image is subjected to road damage detection;

[0035] The improved Yolov5 algorithm is as follows: based on the Yolov5 algorithm, the FReLU activation function is introduced to expand it for image recognition tasks, and a positioning loss function is introduced to locate the target detection frame.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The present invention measures the vibration of the vehicle by setting up a vehicle-mounted vibration detection system, which can improve the acquisition speed and measurement accuracy of the sensor, perform signal denoising according to the characteristics of the vehicle vibration signal collected by the sensor, and identify and distinguish the waveform data generated by different vibrations, so as to quickly and accurately extract abnormal events. Different energy features correspond to different abnormal events, and the energy of the discrete signal is obtained by feature extraction. The IMF normalized energy feature is used as the input of the SVM classifier, which can exclude the features of non-road damage such as speed bumps, thereby accurately identifying the road damage and eliminating the interference factors;

[0038] 2. The present invention first reflects the vibration condition of the object being measured by measuring the amplitude of the relative vibration. The entire detection process is completed during the normal driving of the vehicle, which does not affect the normal driving and passage of the vehicle, and does not require professional personnel to conduct on-site measurements of the road. It can automatically measure, transmit and infer damage results, and then preliminarily and quickly find the situation of road damage. Then, when the situation of road damage is extracted, it can capture images of existing targets, and can further confirm and analyze the degree of damage and determine the type of damage based on the preliminary location of the damage, so as to inform road maintenance personnel in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of the overall method flow provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figure 1 The present invention provides a technical solution: a method for rapid detection of vehicle-mounted road surface damage based on machine vision, comprising the following steps:

[0042] S1. Collecting vibration data: Install the vehicle-mounted vibration detection system on the vehicle. When the vehicle encounters a pothole road during driving, the vehicle body will vibrate. The vehicle vibration is automatically measured during driving. The vehicle-mounted vibration detection system automatically sends the vibration data and geographic location information to the terminal software through the wireless transmission network, and the terminal software processes the data and determines the road damage;

[0043] S2. Processing vibration data: Input a section of vibration data, first perform maximum point filtering to remove abnormal point data that affects the average value judgment, then perform trend item filtering and empirical mode decomposition (EMD) denoising to filter out trend changes caused by small displacements of the vehicle-mounted vibration detection system and filter out high-frequency small signal background noise;

[0044] S3, extracting abnormal events: the input vibration data has different vibration types, including normal events, abnormal events and interference events. The interference events for abnormal event identification are filtered out by the average value method, and the abnormal events are extracted from the denoised data by the mean square value method;

[0045] S4. Classification of abnormal events: Different categories of abnormal events have different signal energies. The features of abnormal events are extracted by extracting IMF energy features, which are used as inputs of the SVM classifier to classify abnormal events, thereby eliminating vibrations caused by non-road damage in abnormal events;

[0046] S5. Collecting road images: A camera is installed on the vehicle. When an abnormal event caused by road damage is identified, the camera is triggered to collect road images. The collected images are grayed and enhanced. After that, the input images are converted into feature maps using a convolutional neural network. After the feature maps are extracted and sampled by multiple convolutional layers and pooling layers, the fully connected layer flattens the extracted final feature maps for target classification operations.

[0047] S6. Identify the type of road damage: Use the improved Yolov5 algorithm to detect the input feature map. During the detection, a large number of target detection frames are generated. When there is a road damage target, the image of the damaged part of the road is cut out, the degree of road damage is analyzed, the type of road damage is identified, and the analysis results are transmitted to the central detection part.

[0048] In step S1, the vehicle-mounted vibration detection system is as follows: a vibration node is provided in the vehicle body, the vibration node includes an inductive sensor and a mass block connected to a spring, the vibration node is fixed on the vehicle body, when the vehicle body is used as the vibrating object to be measured, the space inside the vehicle is regarded as a reference system for vibration measurement, the vehicle body vibrates, the inductive sensor and the vehicle body are relatively stationary, and the spring drives the mass block to generate vibration different from the object to be measured, relative vibration is generated between the probe of the inductive sensor and the mass block, and the vibration condition of the object to be measured is reflected by measuring the amplitude of the relative vibration;

[0049] The entire measurement process includes a vehicle-mounted part and a central detection part. The vehicle-mounted part is used to detect, collect, and transmit vibration data and geographic location information, while the central detection part is used to receive data from the vehicle-mounted vibration detection system and identify the type of road damage.

[0050] Step S2 includes the following steps:

[0051] S201, the maximum value point filtering process is to segment a section of noisy data, find the maximum value point of each section of data, and calculate the average value of each section of data, and replace the found maximum value point with the average value;

[0052] S202, using the least squares method to filter out the trend term, the empirical mode decomposition (EMD) denoising is to use the EMD signal decomposition algorithm to perform time-frequency decomposition from the local characteristics of the signal itself, obtain the results of different frequency decompositions, and construct a low-pass filter based on the results to filter the background noise of the remaining high-frequency small signal in the signal;

[0053] In step S202, the EMD signal decomposition algorithm is specifically:

[0054]

[0055] Among them, X(t) represents the original signal, IMF i (t) represents the IMF component. Each of the n IMF components represents a signal component of a frequency in the original signal. n represents the total number of IMF components. n (t) represents the residual component, t represents time, i represents the IMF component number, i∈[1,n];

[0056] The low-pass filter construction process is as follows: decompose the noisy signal to obtain the IMF and residual components, calculate the signal energy of each IMF component and the residual component, then find the minimum value of the IMF component signal energy, add one to the minimum value as the demarcation point, select the demarcation point and the IMF components after it to reconstruct the denoised signal;

[0057] In step S3, a normal event refers to a waveform generated when a vehicle runs over a smooth and healthy road surface; an interference event refers to a waveform generated when a vibration node is detected to be vibrating without the vehicle vibrating, including vehicle speed change, braking and turning; an abnormal event refers to a waveform generated when a vehicle runs over a damaged road surface and a waveform generated by obvious vibration such as a speed bump.

[0058] In step S3, when an interference event occurs, the mass block in the vibration node generates an extremely low-frequency change in the equilibrium position, and the data generated by the vibration node deviating from the equilibrium position is obtained by average value calculation, and the data whose absolute value of the average value deviation exceeds the set threshold is determined as an interference event, and the data whose absolute value does not exceed the set threshold is determined as an abnormal event and a normal event;

[0059] The process of selecting the average value threshold is as follows: select three types of signal waveform data, namely normal events, interference events and abnormal events, for average value analysis, and calculate the absolute value of the average value deviation of each event. The absolute value of the average value deviation refers to the difference between the true value of the signal and the average value, and obtain the result of statistical analysis of the selected data. The threshold is set according to the result of statistical analysis.

[0060] Calculate the mean square value of each data segment, and determine the data whose mean square value deviation exceeds the set threshold as an abnormal event, otherwise it is a normal event;

[0061] The process of selecting the mean square value threshold is as follows: selecting two types of signal waveform data, normal events and abnormal events, for mean square value analysis, and calculating the mean square value deviation of each event to obtain the result of statistical analysis of the selected data, and setting the threshold according to the result of statistical analysis;

[0062] In step S4, the abnormal event data is decomposed by the EMD signal decomposition algorithm in step S2 to obtain n IMF components and residual components, and the ratio of the energy of each IMF component to the signal energy is used as the feature vector. The specific steps are as follows:

[0063] S401, calculate the energy E of each time domain IMF component i ;

[0064]

[0065] Among them, E i represents the energy of each time domain IMF component, c i (t) represents the energy of the IMF component of a single time domain, N represents the total time domain, t represents the time, i represents the time sequence number, n represents the last time, and I represents the time sequence number;

[0066] S402, constructing a vector T with the energy of each time-domain IMF component, and obtaining an IMF normalized energy feature T' through normalization processing;

[0067] T=[E1 E2···E n ];

[0068]

[0069] Where T represents the energy vector of each time domain IMF component, T' represents the IMF normalized energy feature, E represents the signal energy, and E i Represents the total energy of each time domain IMF component, i represents the numerical sequence, i∈[1,n], n represents the total time domain;

[0070] In step S6, the target detection frame includes the position of the target detection frame and the confidence information of the target detection frame, and the target detection frame with low confidence is eliminated by using the non-maximum suppression method, and then the road damage detection is performed on the acquired road surface image;

[0071] The non-maximum suppression method is to suppress overlapping or similar target detection frames after obtaining a series of target detection frames and their corresponding confidence scores, ensuring that only one most accurate target detection frame is retained for the same object, thereby reducing false positives and repeated detections and improving the accuracy of target detection.

[0072] The improved Yolov5 algorithm is as follows: based on the Yolov5 algorithm, the FReLU activation function is introduced to expand it for image recognition tasks, and a positioning loss function is introduced to locate the target detection box;

[0073] The FReLU activation function is:

[0074] f(x c,i,j )=max(x c,i,j ,T(x c,i,j ));

[0075]

[0076] Among them, f(x c,i,j ) represents the FReLU activation function, xc,i,j represents the pixel point in the i-th row and j-th column on the c-th channel of the input feature map, T(x c,i,j ) represents the parameter, max represents the maximum value function, represents a parameter pooling window of size h×w centered at pixel xc,i,j. Represents the coefficients shared on the same channel on the parameter pooling window;

[0077] The positioning loss function is extended based on the IOU loss function, specifically:

[0078] L = 1-IOU + αβ;

[0079]

[0080] Among them, L represents the positioning loss function, α and β represent parameters, IOU represents the calculation formula of IOU loss function, π represents pi, and w gt and h gt They represent the width and height of the real target box, respectively, and w and h represent the width and height of the predicted target box.

[0081] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0082] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for rapid detection of vehicle-mounted road surface damage based on machine vision, characterized in that: The method comprises the following steps: S1. Collecting vibration data: Install the vehicle-mounted vibration detection system on the vehicle. When the vehicle encounters a pothole road during driving, the vehicle body will vibrate. The vehicle vibration is automatically measured during driving. The vehicle-mounted vibration detection system automatically sends the vibration data and geographic location information to the terminal software through the wireless transmission network, and the terminal software processes the data and determines the road damage; S2. Processing vibration data: Input a section of vibration data, first perform maximum point filtering to remove abnormal point data that affects the average value judgment, then perform trend item filtering and empirical mode decomposition (EMD) denoising to filter out trend changes caused by small displacements of the vehicle-mounted vibration detection system and filter out high-frequency small signal background noise; S3, extracting abnormal events: the input vibration data has different vibration types, including normal events, abnormal events and interference events. The interference events for abnormal event identification are filtered out by the average value method, and the abnormal events are extracted from the denoised data by the mean square value method; S4. Classification of abnormal events: Different categories of abnormal events have different signal energies. The features of abnormal events are extracted by extracting IMF energy features, which are used as inputs of the SVM classifier to classify abnormal events, thereby eliminating vibrations caused by non-road damage in abnormal events; S5. Collecting road images: A camera is installed on the vehicle. When an abnormal event caused by road damage is identified, the camera is triggered to collect road images. The collected images are grayed and enhanced. After that, the input images are converted into feature maps using a convolutional neural network. After the feature maps are extracted and sampled by multiple convolutional layers and pooling layers, the fully connected layer flattens the extracted final feature maps for target classification operations. S6. Identify the type of road damage: Use the improved Yolov5 algorithm to detect the input feature map. During the detection, a large number of target detection frames are generated. When there is a road damage target, the image of the damaged part of the road is cut out, the degree of road damage is analyzed, the type of road damage is identified, and the analysis results are transmitted to the central detection part.

2. The method for rapid detection of vehicle-mounted road surface damage based on machine vision according to claim 1, characterized in that: In the step S1, the vehicle-mounted vibration detection system is as follows: a vibration node is provided in the vehicle body, the vibration node includes an inductive sensor and a mass block connected to a spring, the vibration node is fixed on the vehicle body, when the vehicle body is used as the vibrating object to be measured, the space inside the vehicle is regarded as a reference system for vibration measurement, the vehicle body vibrates, the inductive sensor and the vehicle body are relatively stationary, and the spring drives the mass block to generate vibration different from the object to be measured, relative vibration is generated between the probe of the inductive sensor and the mass block, and the vibration condition of the object to be measured is reflected by measuring the amplitude of the relative vibration; The entire measurement process includes a vehicle-mounted part and a central detection part. The vehicle-mounted part is used to detect, collect, and transmit vibration data and geographic location information, while the central detection part is used to receive data from the vehicle-mounted vibration detection system and identify the type of road damage.

3. The method for rapid detection of road surface damage based on vehicle-mounted machine vision according to claim 1, characterized in that: The step S2 comprises the following steps: S201, the maximum value point filtering process is to segment a section of noisy data, find the maximum value point of each section of data, and calculate the average value of each section of data, and replace the found maximum value point with the average value; S202. Use the least squares method to filter out the trend term. Empirical mode decomposition (EMD) denoising uses the EMD signal decomposition algorithm to perform time-frequency decomposition based on the local characteristics of the signal itself to obtain results of decomposition at different frequencies. A low-pass filter is constructed based on the results to filter the background noise of the remaining high-frequency small signals in the signal.

4. The method for rapid detection of vehicle-mounted road surface damage based on machine vision according to claim 3 is characterized in that: In step S202, the EMD signal decomposition algorithm is specifically: Among them, X(t) represents the original signal, IMF i (t) represents the IMF component. Each of the n IMF components represents a signal component of a frequency in the original signal. n represents the total number of IMF components. n (t) represents the residual component, t represents time, i represents the IMF component number, i∈[1,n]; The low-pass filter construction process is as follows: decomposing the noisy signal to obtain IMF components and residual components, calculating the signal energy of each IMF component and the residual component, and then finding the minimum value of the IMF component signal energy. The minimum value is incremented by one as the dividing point, and the IMF components at and after the dividing point are selected for reconstruction to obtain the denoised signal.

5. The method for rapid detection of road surface damage based on vehicle-mounted machine vision according to claim 1, characterized in that: In step S3, the normal event refers to the waveform generated by the vehicle running over a flat and healthy road surface, the interference event refers to the waveform generated when the vehicle does not vibrate but the vibration node is detected to be shaking, including vehicle speed change, braking and turning, and the abnormal event refers to the waveform generated by the vehicle running over a damaged road surface and the obvious vibration such as running over artificial obstacles such as speed bumps.

6. The method for rapid detection of vehicle-mounted road surface damage based on machine vision according to claim 5, characterized in that: In step S3, when an interference event occurs, the mass block in the vibration node generates an extremely low-frequency change in the equilibrium position, and the data generated by the vibration node deviating from the equilibrium position is obtained by average value calculation, and the data whose absolute value of the average value deviation exceeds the set threshold is determined as an interference event, and the data whose absolute value does not exceed the set threshold is determined as an abnormal event and a normal event; The process of selecting the average value threshold is as follows: select three types of signal waveform data, namely normal events, interference events and abnormal events, for average value analysis, and calculate the absolute value of the average value deviation of each event. The absolute value of the average value deviation refers to the difference between the true value of the signal and the average value, and obtain the result of statistical analysis of the selected data. The threshold is set according to the result of statistical analysis. Calculate the mean square value of each data segment, and determine the data whose mean square value deviation exceeds the set threshold as an abnormal event, otherwise it is a normal event; The mean square value threshold selection process is as follows: two types of signal waveform data, normal events and abnormal events, are selected for mean square value analysis, and the mean square value deviation of each event is calculated to obtain the results of statistical analysis of the selected data, and the threshold is set according to the results of the statistical analysis.

7. The method for rapid detection of vehicle-mounted road surface damage based on machine vision according to claim 1, characterized in that: In step S4, the abnormal event data is decomposed by the EMD signal decomposition algorithm in step S2 to obtain n IMF components and residual components, and the ratio of the energy of each IMF component to the signal energy is used as the feature vector. The specific steps are as follows: S401, calculate the energy E of each time domain IMF component i ; Among them, E i represents the energy of each time domain IMF component, c i (t) represents the energy of the IMF component of a single time domain, N represents the total time domain, t represents the time, i represents the time sequence number, n represents the last time, and I represents the time sequence number; S402, constructing a vector T with the energy of each time-domain IMF component, and obtaining an IMF normalized energy feature T' through normalization processing; Where T represents the energy vector of each time domain IMF component, T' represents the IMF normalized energy feature, E represents the signal energy, and E i Represents the total energy of each time domain IMF component, i represents the numerical sequence, i∈[1,n], and n represents the total time domain.

8. The method for rapid detection of road damage based on vehicle-mounted machine vision according to claim 1, characterized in that: In step S6, the target detection frame includes the position of the target detection frame and the confidence information of the target detection frame, and the target detection frame with low confidence is eliminated by using the non-maximum suppression method, and then the acquired road surface image is subjected to road damage detection; The improved Yolov5 algorithm is as follows: based on the Yolov5 algorithm, the FReLU activation function is introduced to expand it for image recognition tasks, and a positioning loss function is introduced to locate the target detection frame.

Images