Artificial intelligence-based unmanned road surface instability factor perception method and system

Abstract

The application discloses an unmanned road surface instability factor perception method and system based on artificial intelligence, relates to the field of artificial intelligence, and comprises the following steps: acquiring relevant parameters of target vehicles around an unmanned vehicle in real time; when lateral slip speed and instantaneous vehicle head time distance meet conditions, a steering wheel angle micro-motion sequence in a previous time window at a current time is extracted; steering wheel zero speed interval proportion and steering angle sign reversal frequency are calculated; an output of a jamming intention prediction model is used to output an intention confidence of a driver's premeditated jamming; a dynamic risk coefficient representing invasion urgency is acquired; a normalized game risk potential is obtained by performing attenuation correction; the normalized game risk potential is compared with an action threshold value; and the unmanned vehicle is controlled to perform a graded braking operation or to exit the graded braking operation in advance. The application solves the problem that the prior art lacks early perception of a driver's intention and style, leading to delayed risk response and difficulty in ensuring driving safety.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an artificial intelligence-based method and system for perceiving road instability factors in autonomous driving. Background Technology

[0002] With the continuous development of intelligent connected vehicle technology, the application of autonomous vehicles on urban open roads is progressing steadily. However, the dynamic uncertainties brought about by other human-driven vehicles in the road environment remain one of the core issues affecting the safety of autonomous driving. Specifically, the illegal cutting in front of a vehicle from an adjacent lane can significantly reduce the safe driving distance in a short period. Especially when road surface conditions are poor, emergency braking can easily lead to loss of control or rear-end collisions, making it a typical example of dynamic instability on the road surface.

[0003] Most existing unstable factor perception solutions are designed to detect the risk of following another vehicle after it has already entered the lane. They cannot identify the driver's premeditated intention to cut in before the behavior is completed. They can only trigger a response after the vehicle has entered the lane, leaving insufficient reaction time for autonomous vehicles. This makes it difficult to cope with sudden risks on low-adhesion road surfaces such as wet, slippery, or icy surfaces. Therefore, there is an urgent need for a perception method that can detect such road instability factors in advance and dynamically adjust the risk threshold according to different road adhesion conditions and vehicle speed to improve the driving safety of autonomous vehicles in open road scenarios. Summary of the Invention

[0004] This application provides an artificial intelligence-based method and system for perceiving road instability factors in autonomous driving, which solves the technical problem that the prior art lacks advance perception of the driver's intentions and style, resulting in delayed risk response and difficulty in ensuring driving safety.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows: In a first aspect, this application provides an artificial intelligence-based method for perceiving road instability factors in autonomous driving, the method comprising: The system acquires in real time the steering wheel angle micro-motion sequence of target vehicles around the autonomous vehicle, the instantaneous head-to-head distance between the autonomous vehicle's front and the target vehicle's rear bumper, and the lateral slip speed of the target vehicle relative to the lane centerline. When the lateral slip velocity is detected to be greater than zero and the instantaneous headway is continuously reduced beyond the dynamic headway compression rate threshold, the steering wheel angle micro-motion sequence within the previous time window at the current moment is extracted. Based on the steering wheel angle micro-motion sequence, the proportion of the steering wheel zero speed range and the frequency of the steering wheel sign reversal are calculated, and then input into the lane-cutting intention prediction model to output the confidence level of the driver's intention to cut in. Based on the intent confidence and the instantaneous headway, and combined with the lateral acceleration of the target vehicle, a dynamic risk coefficient characterizing the urgency of the intrusion is obtained. Based on the adhesion coefficient of the road surface where the autonomous vehicle is currently located and the current speed of the vehicle, the dynamic risk coefficient is attenuated and corrected to obtain the normalized game risk potential energy. When the normalized game risk potential energy exceeds the first action threshold, the driverless vehicle is controlled to perform a graded braking operation with deceleration increasing over time. When the normalized game risk potential energy falls back below the second action threshold, the graded braking operation is prematurely terminated.

[0006] Secondly, this application provides an artificial intelligence-based autonomous driving road instability factor perception system, including: The vehicle parameter acquisition module is used to acquire in real time the steering wheel angle micro-motion sequence of target vehicles around the autonomous vehicle, the instantaneous head-to-head distance between the front of the autonomous vehicle and the rear bumper of the target vehicle, and the lateral slip speed of the target vehicle relative to the center line of the lane. The steering wheel angle extraction module is used to extract the steering wheel angle micro-motion sequence within the previous time window when the lateral slip speed is detected to be greater than zero and the instantaneous headway is continuously reduced beyond the dynamic headway compression rate threshold. The lane-cutting intention prediction module is used to calculate the proportion of the zero-speed range of the steering wheel and the reversal frequency of the steering wheel sign based on the steering wheel angle micro-motion sequence, and input them into the lane-cutting intention prediction model to output the confidence level of the driver's intention to cut in. The dynamic risk coefficient acquisition module is used to acquire a dynamic risk coefficient representing the urgency of intrusion based on the intent confidence and the instantaneous headway of the vehicle, and in combination with the lateral acceleration of the target vehicle. The dynamic risk coefficient correction module is used to attenuate and correct the dynamic risk coefficient based on the adhesion coefficient of the road surface where the unmanned vehicle is currently located and the current speed of the vehicle, so as to obtain the normalized game risk potential energy. The graded braking execution module is used to control the autonomous vehicle to perform graded braking operations with increasing deceleration over time when the normalized game risk potential energy exceeds the first action threshold, and to exit the graded braking operation in advance when the normalized game risk potential energy falls back below the second action threshold.

[0007] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides an AI-based method and system for perceiving road instability factors in autonomous driving. First, by extracting the micro-motion characteristics of the target vehicle's steering wheel angle, it identifies premeditated cutting-in intentions before the act is completed, only during the driver's initial fine-tuning preparation stage. Second, by combining the target vehicle's instantaneous headway and lateral acceleration to obtain a dynamic risk coefficient, and then combining the current road surface adhesion conditions and the vehicle's current speed to correct the dynamic risk coefficient, a normalized game-theoretic risk potential energy matching the current driving scenario is obtained, adaptable to the risk response requirements of different road surfaces. Finally, a graded, incremental braking control strategy and an early exit mechanism are adopted to avoid the risk of vehicle skidding and loss of control caused by emergency braking, and also to avoid unnecessary continuous braking affecting traffic efficiency, effectively improving the driving safety of autonomous vehicles dealing with dynamic road instability factors on open roads.

[0008] Through the above technical solution, this application solves the technical problem that the existing technology cannot identify the driver's premeditated intention to cut in front of the vehicle before the cutting behavior is completed, and can only trigger the response after the vehicle enters the lane. This leaves insufficient reaction time for the autonomous vehicle and makes it difficult to cope with sudden risks under different road conditions. The new technology can perceive the dynamic instability factors of the road surface in advance, match different road surface adhesion conditions and vehicle speed to dynamically adjust the risk threshold and braking strategy, take into account driving safety and traffic efficiency, and adapt to the driving needs of various complex road surface scenarios in open roads. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the artificial intelligence-based method for perceiving road instability factors in autonomous driving, as provided in an embodiment of this application. Figure 2 This is a flowchart illustrating the calculation of dynamic risk coefficients in the AI-based unmanned driving road instability factor perception method provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an artificial intelligence-based autonomous driving road instability factor perception system provided in an embodiment of this application.

[0011] The components represented by each number in the attached diagram are explained below: Vehicle parameter acquisition module 11, steering wheel angle extraction module 12, lane-cutting intention prediction module 13, dynamic risk coefficient acquisition module 14, dynamic risk coefficient correction module 15, graded braking execution module 16. Detailed Implementation

[0012] This application provides an artificial intelligence-based method and system for perceiving road instability factors in autonomous driving, which addresses the technical problem that existing technologies lack advance perception of driver intentions and styles, leading to delayed risk response and difficulty in ensuring driving safety.

[0013] Example 1, as Figure 1 As shown in the embodiments of this application, an artificial intelligence-based method for perceiving road instability factors in autonomous driving is provided, including: S10: Real-time acquisition of the steering wheel angle micro-motion sequence of target vehicles around the autonomous vehicle, the instantaneous head-to-head distance between the front of the autonomous vehicle and the rear bumper of the target vehicle, and the lateral slip speed of the target vehicle relative to the lane centerline. In this embodiment, relevant parameters of target vehicles around the autonomous vehicle are acquired in real time. Specifically, the steering wheel angle micro-motion data of target vehicles in adjacent lanes can be acquired through the vehicle-mounted camera and arranged in chronological order to form a steering wheel angle micro-motion sequence. The distance between the front of the autonomous vehicle and the rear bumper of the target vehicle is acquired through the millimeter-wave radar and camera fusion perception scheme on the autonomous vehicle. The instantaneous front-to-rear distance is calculated by combining the current speed of the autonomous vehicle. At the same time, the lateral offset of the target vehicle in the direction perpendicular to the center line of the lane is acquired. The lateral slip speed of the target vehicle relative to the center line of the lane is calculated by combining the sampling time interval.

[0014] Specifically, step S10 in the method includes: The vehicle continuously acquires images of the front tire treads of the target vehicle using an onboard camera, and extracts estimated values ​​of the front wheel steering angle from the images of the front tire treads. The estimated value of the front wheel steering angle is recorded at a fixed sampling period, and combined with the steering system angular transmission ratio to generate the instantaneous value of the steering wheel angle. The steering system angular transmission ratio is the ratio of the maximum rotation angle of the target vehicle's steering wheel to the maximum mechanical deflection angle of the front wheels. The instantaneous values ​​of steering wheel angle obtained from N consecutive sampling periods are arranged in chronological order to form the steering wheel angle micro-motion sequence, where N is an integer between 50 and 100; The longitudinal relative distance between the front of the autonomous vehicle and the rear bumper of the target vehicle is measured synchronously by millimeter-wave radar, the current speed of the autonomous vehicle is read, and the instantaneous front-to-rear distance is calculated. The vehicle-mounted camera identifies the position of the lane centerline, detects the lateral offset of the midpoint of the target vehicle in the direction perpendicular to the lane centerline, and divides the difference of the lateral offset obtained in two consecutive sampling periods by the sampling time interval to obtain the lateral slip velocity of the target vehicle relative to the lane centerline. The instantaneous values ​​of steering wheel angle, instantaneous headway, and lateral slip speed collected at the same sampling moment are aligned by timestamps to form an input parameter set synchronized with the time window of the steering wheel angle micro-motion sequence.

[0015] In this embodiment, firstly, the front tire tread images of the target vehicle are continuously acquired by the vehicle-mounted camera, and the contour features of the tread pattern and the wheel hub edge are extracted. Based on the relative position of the contours, the deflection angle of the front tire relative to the longitudinal axis of the vehicle body is calculated, thereby obtaining the estimated value of the front wheel steering angle.

[0016] Secondly, the estimated front wheel steering angle is recorded according to a preset fixed sampling period. For example, the preset fixed sampling period can be set to 10ms to 20ms. Combined with the preset steering system angular transmission ratio of the target vehicle model, the corresponding instantaneous value of the steering wheel angle is calculated, that is, the instantaneous value of the steering wheel angle = the estimated value of the front wheel steering angle × (maximum steering wheel rotation angle / maximum front wheel mechanical deflection angle), where the maximum steering wheel rotation angle is 540 degrees and the maximum front wheel mechanical deflection angle is 30 degrees.

[0017] Next, the instantaneous values ​​of the steering wheel angle obtained from N consecutive sampling periods are arranged in chronological order to form a steering wheel angle micro-motion sequence for feature extraction. For example, in this embodiment, N can be 80, which can cover the driver's steering adjustment actions within 3 seconds. This will not introduce irrelevant data due to an excessively large time window, nor will it miss key micro-motion features due to an excessively small time window.

[0018] Then, the target vehicle is continuously measured by millimeter-wave radar, and the current speed of the autonomous vehicle is read simultaneously. The measured longitudinal relative distance is divided by the current speed of the autonomous vehicle to obtain the instantaneous headway, that is, the instantaneous headway = the longitudinal relative distance between the front of the autonomous vehicle and the rear bumper of the target vehicle / the current speed of the autonomous vehicle. This instantaneous headway reflects the longitudinal safety margin between the two vehicles.

[0019] Next, the vehicle's camera identifies the center lines of the two lanes on the current road, detects the coordinates of the target vehicle's midpoint, and calculates the vertical distance from the midpoint to the center line of the vehicle's lane, which is the lateral offset at the current moment. Then, the difference between the lateral offsets of two adjacent sampling periods is divided by the sampling time interval to obtain the lateral slip velocity of the target vehicle relative to the lane center line, i.e., lateral slip velocity = (lateral offset of the next sampling period - lateral offset of the previous sampling period) / sampling time interval. A lateral slip velocity greater than zero indicates that the target vehicle is moving towards the center line of the vehicle's lane, while a velocity less than zero indicates that the target vehicle is moving away from the vehicle's lane.

[0020] Finally, the instantaneous values ​​of steering wheel angle, instantaneous headway, and lateral slip speed collected at the same sampling moment are aligned by timestamps to ensure that the time windows of all input parameters remain synchronized. The specific alignment method can be achieved by linear interpolation to compensate for the time deviation caused by the difference in sampling frequency of different sensors, ensuring that all data can correspond to the same time node and avoiding feature extraction errors due to time misalignment. This forms an input parameter set synchronized with the time window of the steering wheel angle micro-motion sequence. The input parameter set is used to determine the triggering condition that the lateral slip speed is greater than zero and the instantaneous headway continuously decreases.

[0021] Specifically, the method involves continuously acquiring images of the front tire treads of the target vehicle using an onboard camera, and extracting an estimated front wheel steering angle from the images, including: Identify the boundary line between the wheel hub edge and the tire tread contact area from the front wheel tire tread image, extract at least two feature points at different radial positions on the boundary line, and connect the feature points to form a wheel hub reference line; Using the horizontal axis of the front tire tread image as the reference direction of the vehicle's longitudinal axis, calculate the angle between the wheel hub reference line and the horizontal axis of the image, and use the angle as the initial front wheel steering angle; Extract the hub center point and the tire tread pattern symmetry axis from the front tire tread image, and calculate the offset angle of the symmetry axis relative to the hub center point and the direction perpendicular to the horizontal axis of the image, as the steering correction angle; The estimated value of the front wheel steering angle is obtained by subtracting the steering correction angle from the initial front wheel steering angle.

[0022] In this embodiment, firstly, the boundary line between the wheel hub edge and the tire tread contact area is identified from the front wheel tire tread image. Since the wheel hub itself is a rigid circular structure, the line connecting its center can reflect the actual steering angle of the front wheel. Therefore, at least two feature points at different radial positions are extracted on the boundary line, and the feature points are connected to form a wheel hub reference line.

[0023] Secondly, the horizontal axis of the front tire tread image is used as the reference direction of the vehicle's longitudinal axis, corresponding to the vehicle's forward direction in the actual driving scenario. The angle between the wheel hub reference line and the horizontal axis of the image is calculated, and this angle is used as the initial front wheel rotation angle. That is, the initial front wheel rotation angle = the difference between the angle between the wheel hub reference line and the horizontal axis of the image. When the wheel hub reference line shifts to the left of the vehicle, the difference is positive, and when it shifts to the right, it is negative.

[0024] Subsequently, the center point of the wheel hub is extracted from the image, and the axis of symmetry of the tire tread pattern is identified. The offset angle of this axis of symmetry relative to the wheel hub center point and the direction perpendicular to the horizontal axis of the image is calculated. Since there is a slight error in tire installation, this offset angle can be used as a steering correction angle to offset the estimation error caused by the installation deviation. Specifically, the steering correction angle is the difference between the angle between the axis of symmetry and the vertical reference line of the wheel hub center point. When the axis of symmetry is offset to the left, the difference is positive, and when it is offset to the right, it is negative.

[0025] Finally, subtracting the steering correction angle from the initial front wheel angle yields a more accurate estimate of the front wheel steering angle, i.e., the estimated front wheel steering angle = initial front wheel angle - steering correction angle, providing accurate basic data for the subsequent generation of the steering wheel angle micro-motion sequence.

[0026] Based on the above step S10, it can be detected whether the triggering conditions for lane-cutting intention recognition are met: that is, whether the acquired lateral slip speed is greater than zero, and whether the instantaneous headway is continuously reduced at a rate exceeding the dynamic headway compression rate threshold. When both conditions are met, the subsequent lane-cutting intention recognition process is triggered; otherwise, the vehicle parameters of the next time window are re-acquired, and the driving status of the target vehicle is monitored.

[0027] S20: When the lateral slip velocity is detected to be greater than zero and the instantaneous headway is continuously reduced beyond the dynamic headway compression rate threshold, extract the steering wheel angle micro-motion sequence within the previous time window at the current moment; In this embodiment of the application, when the target vehicle simultaneously meets two conditions—lateral sliding toward the lane and rapid compression of the longitudinal distance with the vehicle—that is, when the lateral sliding speed is greater than zero and the instantaneous headway is reduced three or more times consecutively exceeding the dynamic headway compression rate threshold, it indicates that the target vehicle has a high potential tendency to cut in line. At this time, the steering wheel angle micro-motion sequence within the previous complete time window is extracted. This steering wheel angle micro-motion sequence covers all steering micro-adjustment actions of the target vehicle driver during the cut-in preparation stage, and is used to subsequently extract steering micro-motion features to identify the cut-in intention.

[0028] The calculation of the dynamic time-distance compression ratio threshold includes: Continuous sampling points whose absolute values ​​of the instantaneous steering wheel angle are less than the steering angle silence threshold are extracted from the steering wheel angle micro-motion sequence to form a silence segment. The steering angle silence threshold is calibrated based on the steering wheel angle sensor noise level and the steering device free travel. The instantaneous headway distance of each sampling point within the silent segment is statistically analyzed, and the average attenuation slope of the instantaneous headway distance within the silent segment is calculated. Combined with the time distance compression amplification factor, the basic compression ratio threshold is calculated. The calculation steps for the time-distance compression amplification factor include: The attenuation difference is obtained by subtracting the instantaneous headway from the end of the sampling point from the instantaneous headway at the beginning of the silent segment. The relative attenuation amplitude is determined based on the attenuation difference and the instantaneous headway at the starting sampling point; Add 1 to the relative attenuation amplitude and use it as the time-distance compression amplification factor; The change in instantaneous value of steering wheel angle between two adjacent sampling points is extracted from the steering wheel angle micro-motion sequence, and the rate of change of angle is calculated by combining the sampling time interval. The duration of all corner change rates exceeding the rate threshold within the time window is summed, and the proportion of aggressive operations is calculated by combining the total duration of the time window. The rate threshold is a calibrated rate threshold that distinguishes between aggressive and normal operations based on historical driving data statistics. Collect data from a large-scale natural driving database, including the rate of change of steering wheel angle of a large number of drivers in a stable following state, and calculate the proportion of aggressive operations in each time window; The percentages of aggressive operations were sorted from smallest to largest, and the 85th percentile was taken as the upper limit of the statistics. The dynamic time-distance compression ratio threshold is determined based on the base compression ratio threshold, the proportion of aggressive operations, and the statistical upper limit value, which are calculated continuously in real time. The base compression ratio threshold and the proportion of aggressive operations are updated online according to the steering wheel angle micro-motion sequence maintained by continuous sliding window. The dynamic time-distance compression ratio threshold is directly proportional to the base compression ratio threshold and the proportion of aggressive operations, and inversely proportional to the statistical upper limit value.

[0029] In this embodiment, firstly, from the steering wheel angle micro-motion sequence composed of several consecutive sampling points, sampling points whose instantaneous absolute value of the steering wheel angle does not exceed the angle silence threshold are selected. The consecutive sampling points are then combined into a silence segment. The angle silence threshold is calibrated in advance based on the inherent noise level of the steering wheel angle sensor and the empty travel range of the vehicle steering device itself. This can filter out minute data fluctuations when there is no steering intention and avoid false triggering of the recognition process.

[0030] Secondly, the instantaneous headway distance corresponding to each sampling point within the silent segment is statistically analyzed. A linear fitting method is used to calculate the average attenuation slope of the instantaneous headway distance throughout the entire silent segment. For example, the average attenuation slope = (instantaneous headway distance at the beginning of the silent segment - instantaneous headway distance at the end of the silent segment) / the time interval between the starting and ending sampling points. This is then combined with the calculated headway compression amplification factor to obtain the basic compression ratio threshold, i.e., the basic compression ratio threshold = average attenuation slope × headway compression amplification factor. This basic compression ratio threshold will be dynamically adjusted according to the current driving state of the target vehicle to adapt to the current following state of the target vehicle.

[0031] Specifically, when calculating the time-distance compression amplification factor, firstly, the instantaneous headway distance at the beginning and end of the silent segment is subtracted from the instantaneous headway distance at the end of the silent segment to obtain the attenuation difference of the headway distance throughout the entire silent segment. Then, the attenuation difference is divided by the instantaneous headway distance at the beginning of the silent segment to obtain the current relative attenuation amplitude of the target vehicle. Finally, the relative attenuation amplitude is added by 1 to obtain the current time-distance compression amplification factor, i.e., time-distance compression amplification factor = relative attenuation amplitude + 1. The larger the relative attenuation amplitude, the larger the time-distance compression amplification factor, and the higher the corresponding basic compression rate threshold, which conforms to the actual risk change pattern. Furthermore, the steering wheel angle micro-motion sequence within the current time window is traversed, and the change in the instantaneous value of the steering wheel angle between each pair of adjacent sampling points is calculated in turn. This change is divided by the sampling time interval between the two sampling points to obtain the angle change rate within that interval. That is, the angle change rate = the absolute value of the difference between the instantaneous values ​​of the steering wheel angle of adjacent sampling points / sampling time interval.

[0032] Next, each instance of steering angle change rate is checked to see if it exceeds a pre-calibrated rate threshold. The duration of all instances meeting the condition is accumulated, and the total accumulated time is divided by the total duration of the entire time window to obtain the proportion of aggressive operations within the current time window. This rate threshold is based on large-scale historical driving data statistics and can clearly distinguish between aggressive steering operations and normal driving adjustment operations. To obtain a calibration reference that conforms to statistical laws, steering wheel angle change rate data of a large number of drivers in a large-scale natural driving database under conditions where no risk events such as cutting in or sudden braking have occurred are collected in advance. After calculating the proportion of aggressive operations window by window, all proportion values ​​are sorted from smallest to largest, and the 85th percentile after sorting is taken as the statistical upper limit value. This value can cover the operation proportion level of most normal driving scenarios.

[0033] Finally, by combining the real-time updated base compression ratio threshold, the currently calculated aggressive operation percentage, and the pre-calibrated statistical upper limit, the final dynamic time-distance compression ratio threshold is calculated according to the formula, for example, dynamic time-distance compression ratio threshold = base compression ratio threshold × (aggressive operation percentage / statistical upper limit). The base compression ratio threshold and the aggressive operation percentage are updated online with the steering wheel angle micro-motion sequence maintained by continuous window sliding. The dynamic time-distance compression ratio threshold is directly proportional to the base compression ratio threshold and the aggressive operation percentage, and inversely proportional to the statistical upper limit. It can dynamically adjust the trigger threshold according to the driving style of the target vehicle driver, avoid the regular operation of aggressive drivers being misjudged as cutting in, and can also identify the cutting inclination of conservative drivers earlier.

[0034] Based on the above step S20, a complete steering wheel angle micro-motion sequence within a specified time window can be extracted. This sequence completely records all steering micro-adjustment actions of the target vehicle driver before the intention to cut in line is triggered. Next, steering micro-motion features that reflect the driver's intention to cut in line can be extracted based on this sequence, providing feature basis for subsequent intention determination.

[0035] S30: Based on the steering wheel angle micro-motion sequence, calculate the proportion of the steering wheel zero speed range and the steering angle sign reversal frequency, and input them into the lane-cutting intention prediction model to output the confidence level of the driver's intention to cut in line. In this embodiment of the application, when extracting the steering micro-motion features used to identify the intention to cut in, the steering wheel angle micro-motion sequence and the steering sign reversal event within the entire time window are first traversed and statistically analyzed to calculate the proportion of the steering wheel zero speed interval and the steering sign reversal frequency. The proportion of the steering wheel zero speed interval reflects the degree of hesitation of the driver when adjusting the steering, and the steering sign reversal frequency reflects the frequency of the driver's steering adjustment.

[0036] The calculated percentage of the steering wheel's zero-speed range and the frequency of the steering sign reversal are input into a pre-trained model for predicting the intention to cut in line. This model is built on a logistic regression algorithm and can directly output the confidence level of the current target vehicle driver's intention to cut in line. The confidence level ranges from 0 to 1, with a higher value indicating a higher probability of cutting in line.

[0037] Among them, based on the steering wheel angle micro-motion sequence, the calculation of the proportion of the steering wheel zero speed range and the steering angle sign reversal frequency includes: Traverse the steering wheel angle micro-motion sequence, compare the instantaneous value of the steering wheel angle at each sampling point with the angle silence threshold, and mark the zero speed point where the absolute value of the angle is less than the angle silence threshold; Divide consecutive adjacent zero-velocity points within the time window into several zero-velocity segments, and sum the sampling duration spanned by each zero-velocity segment to obtain the total zero-velocity duration. Calculate the ratio of the total zero-speed time to the total time of the time window, and use it as the proportion of the zero-speed range of the steering wheel; Based on the steering wheel angle micro-motion sequence, the sampling points marked as zero speed points are removed to obtain the non-zero speed point sequence; Iterate through the instantaneous steering wheel angle values ​​of two adjacent sampling points in the non-zero speed point sequence, and check whether the signs of the angle values ​​of the two adjacent points are opposite. If the signs are opposite, count one angle sign reversal event. The total number of corner sign reversal events occurring within the time window is accumulated, and the ratio of the total number to the total number of adjacent sampling point pairs in the non-zero velocity point sequence is calculated as the corner sign reversal frequency.

[0038] In this embodiment, firstly, each sampling point in the steering wheel angle micro-motion sequence is traversed, and the absolute value of the instantaneous steering wheel angle of each sampling point is compared with the pre-calibrated steering angle silence threshold. As long as the absolute value of the instantaneous value is less than the threshold, the sampling point is marked as a zero speed point. This step can quickly filter out the sampling points where the driver has not made active steering adjustments, corresponding to the driver's hesitant and undecided operation state of cutting in.

[0039] Secondly, consecutive adjacent zero-speed points within the time window are merged and divided into several independent zero-speed segments. The sampling duration spanned by each zero-speed segment is counted one by one. Then, the durations of all zero-speed segments are added together to obtain the total zero-speed duration within the entire time window. Finally, the total zero-speed duration is divided by the total duration of the entire time window to obtain the proportion of the steering wheel zero-speed interval. The higher the proportion of the zero-speed interval, the more hesitation and pauses the driver makes during steering operations, and the higher the possibility of premeditated lane cutting.

[0040] Next, after calculating the proportion of the zero-speed interval, all sampling points that have been marked as zero-speed points are removed from the original steering wheel angle micro-motion sequence, and all non-zero-speed points are retained to obtain the non-zero-speed point sequence. Then, all pairs of adjacent sampling points in the non-zero-speed point sequence are traversed, and the signs of the instantaneous steering wheel angle values ​​of two adjacent sampling points are checked in turn. Each time a sign is reversed, the counter for the steering wheel angle sign reversal event is incremented by one.

[0041] Finally, the ratio obtained by dividing the total number of corner sign reversal events accumulated within the time window by the total number of adjacent sampling point pairs in the non-zero speed point sequence is the corner sign reversal frequency. The higher the reversal frequency, the more frequently the driver makes steering corrections and the more obvious the intention to cut in line.

[0042] Furthermore, the construction steps of the preemptive cut-in intention prediction model include: Collect multiple instances of target vehicles cutting in front of your vehicle from adjacent lanes during historical driving, as well as an equal number of normal following events; The proportion of the steering wheel's zero-speed range and the frequency of the turning sign reversal within the time window preceding each event are extracted to form a sample feature set. Each sample feature set is labeled with a corresponding label for the actual lane-cutting intention. The label for lane-cutting events is set to 1, and the label for normal following events is set to 0. Based on the logistic regression algorithm, an initial lane-cutting intention prediction model is constructed. The initial lane-cutting intention prediction model uses the proportion of the steering wheel zero speed range and the frequency of the steering angle sign reversal as two input features. The two features are multiplied by their corresponding weights and then added together. The sum is then mapped to the 0 to 1 range by the Sigmoid function, and the output value is used as the confidence level of the driver's intention to cut in. The proportion of the steering wheel zero-speed interval and the frequency of the corner sign reversal in the sample feature set are input into the initial lane-cutting intention prediction model to obtain the prediction intention confidence. The binary cross-entropy loss between the prediction intention confidence and the actual lane-cutting intention label is calculated. With minimizing the binary cross-entropy loss as the training objective, the gradient descent method is used to iteratively optimize the two weight parameters in the initial preemptive intention prediction model. After each iteration, the average cross-entropy loss on the validation set is calculated. Training stops when the average value of the cross-entropy loss decreases by less than the preset convergence threshold for several consecutive iterations, and the current weight parameters are fixed to obtain the trained model for predicting the intention to cut in line.

[0043] In this embodiment, firstly, samples of two types of events, namely, cutting in and normal following, are collected to ensure a balanced number of samples, avoid class bias during model training, and ensure the model's sensitivity in recognizing cutting in events.

[0044] Secondly, following the statistical methods described above, the proportion of zero-speed intervals and the frequency of corner sign reversals within the corresponding time window before the cutting-in behavior occur are extracted for each sample. These are then organized into a structured sample feature set, and each sample is matched with a corresponding ground truth label to indicate whether it represents an actual cutting-in event. The label for a cutting-in event is 1, while the label for a normal following event is 0. Since cutting-in intent recognition is a typical binary classification task involving only two input features, a lightweight logistic regression algorithm is chosen to build the model. This algorithm meets the computational power requirements for real-time recognition and can stably output confidence results in the 0-1 range, achieving prediction without a complex network structure.

[0045] Furthermore, during model training, the binary cross-entropy loss between the predicted result and the true label is continuously reduced through iteration, and the weight parameters corresponding to the two features are gradually optimized. When the decrease in the validation set loss is lower than the preset convergence threshold multiple times in a row, it indicates that the model has converged. Training is stopped and the parameters are fixed. The final model can accurately output the confidence level of the intention to cut in line, which conforms to the actual driving pattern.

[0046] For example, a model for predicting lane-jumping intentions is constructed and trained based on the logistic regression algorithm. The specific steps are as follows: Assume a preset convergence threshold of 1e-5, a requirement of 5 consecutive stop judgments, and initial weight parameters w1 (corresponding to the proportion of the zero-speed interval) and w2 (corresponding to the corner sign reversal frequency) are both initialized to 0.5. The bias b is initialized to 0. All samples are randomly divided into training and validation sets in an 8:2 ratio. Each time, 16 samples are randomly selected from the training set to form a mini-batch. The proportion of the zero-speed interval (x1) and the corner sign reversal frequency (x2) of each sample are input into the model, and the model formula is applied. Calculate the predicted confidence p for each sample. Based on the predicted confidence p and the true label y for all samples in the mini-batch, calculate the average binary cross-entropy loss. , where N is the number of samples in the mini-batch.

[0047] Mini-batch gradient descent with a learning rate of 0.01 is used to calculate the gradients for w1, w2, and b, and update the parameter values. After each round of traversal of all training set samples, the average cross-entropy loss of the current parameters on the validation set is calculated, and the difference between the validation loss of this round and the validation loss of the previous round is recorded. When the decrease in validation loss for 5 consecutive rounds is less than 1e-5, the model is deemed to meet the convergence condition, training is stopped, and the current parameters w1, w2, and b are fixed to obtain the final usable model for predicting lane-jumping intentions.

[0048] Based on the above step S30, the confidence level of the driver's premeditated cutting-in intention in the range of 0 to 1 can be obtained. This confidence level of premeditated cutting-in intention reflects the probability that the driver of the target vehicle currently has a tendency to cut in, providing a quantitative basis for subsequent instability risk assessment.

[0049] S40: Based on the intent confidence and the instantaneous headway, and combined with the lateral acceleration of the target vehicle, obtain a dynamic risk coefficient characterizing the urgency of the intrusion; In this embodiment of the application, after obtaining the confidence level of the driver of the target vehicle's intention to cut in, a dynamic risk coefficient that can characterize the urgency of the current risk is calculated in combination with the current actual motion state of the target vehicle. The lateral acceleration is used to reflect the intensity of the target vehicle's current movement into the lane, and the instantaneous headway is used to reflect the size of the safety gap that the vehicle needs to leave after cutting in.

[0050] Specifically, step S40 in the method includes: The reciprocal of the instantaneous headway is taken as the longitudinal approach rate, and combined with the intention confidence, a weighted longitudinal approximation value is determined. The lateral acceleration of the target vehicle is obtained, and the absolute value of the lateral acceleration is taken as the lateral disturbance intensity. The lateral acceleration is determined based on the ratio of the change in the lateral slip velocity at adjacent sampling times to the sampling time interval. Based on the intent confidence level, horizontal and vertical weights are determined, wherein the horizontal weight decreases as the intent confidence level increases, and the vertical weight increases as the intent confidence level increases. Based on the horizontal weights and the vertical weights, the weighted vertical approximation value and the horizontal interference intensity are weighted and fused to obtain the dynamic risk coefficient.

[0051] In the embodiments of this application, such as Figure 2 As shown, firstly, the longitudinal approach rate is obtained by taking the reciprocal of the instantaneous headway. The smaller the instantaneous headway, the closer the target vehicle is to the vehicle. The larger the longitudinal approach rate, the higher the corresponding risk urgency. Therefore, taking the reciprocal can directly convert distance information into a positive vectorized indicator of risk level.

[0052] Secondly, the longitudinal approach rate is multiplied by the intention confidence to obtain the weighted longitudinal approximation value. The higher the intention confidence, the higher the risk of the weighted longitudinal approximation.

[0053] Subsequently, based on the change in the lateral slip velocity of the target vehicle at adjacent sampling times, the lateral acceleration of the target vehicle is obtained by dividing it by the time interval between adjacent sampling times. That is, lateral acceleration = (lateral slip velocity at the next sampling time - lateral slip velocity at the previous sampling time) / time interval between adjacent sampling times. The absolute value of the lateral acceleration is taken as the lateral interference intensity. The larger the absolute value, the more violent the lateral cutting action of the target vehicle is, and the higher the current risk.

[0054] Furthermore, based on the obtained intent confidence level, corresponding horizontal and vertical weights are matched. A higher intent confidence level indicates a clearer intention from the driver of the target vehicle to cut in line. In this case, the urgency of the vertical distance contributes more to the risk. Therefore, the vertical weight is set to increase as the graph confidence level increases, while the horizontal weight decreases as the graph confidence level increases. Conversely, if the intent confidence level is low, the driver's tendency to cut in line is unclear. In this case, the acceleration of lateral entry better reflects the risk of sudden intrusion. Therefore, the proportion of the horizontal weight is increased, and the proportion of the vertical weight is decreased. For example, the horizontal and vertical weights can be set as follows: Horizontal weight = 1 - Intent confidence level, Vertical weight = Intent confidence level.

[0055] Finally, according to the set lateral and longitudinal weights, the weighted longitudinal approximation value and the lateral interference intensity are weighted and fused to obtain the final dynamic risk coefficient. This coefficient integrates three types of information: the driver's intention to cut in, the target vehicle's position and motion state, and can accurately characterize the urgency of the road intrusion risk posed by the target vehicle to the current vehicle.

[0056] Based on the above step S40, a dynamic risk coefficient can be calculated. The higher the value of the coefficient, the higher the risk of road intrusion and instability caused by the target vehicle to the vehicle. By integrating the driver's subjective intention to cut in with the objective motion state of the target vehicle, the potential risks brought by the target vehicle in the current autonomous driving scenario can be quantified more accurately.

[0057] S50: Based on the adhesion coefficient of the road surface where the unmanned vehicle is currently located and the current speed of the vehicle, the dynamic risk coefficient is attenuated and corrected to obtain the normalized game risk potential energy. In this embodiment, the dynamic risk coefficient only reflects the intrusion risk posed by the target vehicle to the vehicle itself, without considering the impact of the current road conditions and the vehicle's own driving state on its braking and collision avoidance capabilities. Therefore, the dynamic risk coefficient is further modified by combining the road adhesion coefficient and the vehicle's current speed to obtain a normalized game risk potential energy that is more in line with the vehicle's actual coping capabilities.

[0058] Specifically, step S50 in the method includes: Read the adhesion coefficient calibration value of the road surface where the unmanned vehicle is currently located; Read the current speed of the autonomous vehicle and use it as the current speed of the driverless vehicle; Based on the adhesion coefficient calibration value and the current vehicle speed, a vehicle speed-road surface joint factor is determined, wherein the vehicle speed-road surface joint factor is inversely proportional to the adhesion coefficient calibration value and directly proportional to the current vehicle speed; Multiplying the dynamic risk coefficient by the vehicle speed-road surface joint factor yields the normalized game risk potential energy.

[0059] In this embodiment of the application, firstly, after acquiring road surface images and identifying the road surface type through the vehicle-mounted camera, the adhesion coefficient value corresponding to the road surface type is read from the vehicle-mounted memory.

[0060] Secondly, the vehicle's current speed is directly read by the onboard sensors. Based on the friction coefficient calibration value and the vehicle's current speed, the vehicle speed-road joint factor is calculated. The lower the friction coefficient, the worse the road friction conditions, the greater the difficulty for the vehicle to brake and decelerate, and the worse the ability to cope with risks. Therefore, the vehicle speed-road joint factor is inversely proportional to the friction coefficient calibration value. On the other hand, the higher the vehicle's current speed, the shorter the time allowed for adjustment after encountering a sudden risk, and the higher the potential risk. Therefore, the vehicle speed-road joint factor is directly proportional to the vehicle's current speed.

[0061] For example, the vehicle speed-road surface joint factor can be defined as: vehicle speed-road surface joint factor = k * (current vehicle speed / adhesion coefficient calibration value), where k is a preset normalization calibration coefficient used to adjust the calculation result to a preset numerical range.

[0062] Finally, the dynamic risk coefficient obtained in step S40 is multiplied by the calculated vehicle speed-road joint factor to obtain the final normalized game risk potential energy. This potential energy can accurately quantify the degree of road instability risk caused by the current target vehicle cutting in, based on the vehicle's actual collision avoidance capability.

[0063] Based on the above step S50, the final instability risk quantification result after correction based on the vehicle's driving conditions can be obtained. The higher the value, the higher the road intrusion instability risk that the target vehicle currently brings to the vehicle. This completes the perception quantification of the instability factors caused by the target vehicle's cutting-in behavior, providing accurate and reliable risk input for the subsequent collision avoidance decision of the autonomous vehicle.

[0064] S60: When the normalized game risk potential energy exceeds the first action threshold, control the autonomous vehicle to perform a graded braking operation with deceleration increasing over time, and exit the graded braking operation in advance when the normalized game risk potential energy falls back to below the second action threshold.

[0065] In this embodiment, when the normalized game risk potential energy exceeds the preset first action threshold, it indicates that the risk of the target vehicle cutting in has reached the level that requires active intervention to avoid collision. It is necessary to control the vehicle to initiate graded braking to reduce the probability of a collision. If the target vehicle stops cutting in and returns to its original lane during the graded braking process, the normalized game risk potential energy will drop rapidly. When it drops below the second action threshold, the graded braking will be discontinued in advance to avoid unnecessary deceleration that affects traffic efficiency.

[0066] The first action threshold is set to be greater than the second action threshold. By setting the interval between the two thresholds, the jerky feeling of frequent starting and stopping of braking can be avoided when the risk fluctuates around the threshold, thus improving ride comfort. The graded braking is set to a deceleration that increases over time, which can adapt to the process of gradually increasing risk.

[0067] Specifically, step S60 in the method includes: Collect the historical normalized game risk potential energy of the unmanned vehicle in a risk-free following state for multiple consecutive time windows, and calculate the 90th percentile of the historical normalized game risk potential energy as the benchmark risk threshold. Based on the current speed of the vehicle, an acceleration coefficient is obtained, wherein the acceleration coefficient is inversely proportional to the current speed of the vehicle; Based on the growth rate coefficient and the benchmark risk threshold, a first action threshold is obtained; The ratio of the median of the historical normalized game risk potential energy to the benchmark risk threshold is calculated and used as the exit ratio coefficient. Combined with the first action threshold, the second action threshold is determined. When the normalized game risk potential energy calculated in real time exceeds the first action threshold, graded braking is triggered. During the initial braking phase, braking is applied according to the first deceleration slope. After each fixed time interval, the currently used deceleration slope is increased by one step until the maximum deceleration slope is reached. The first deceleration slope is determined based on the current road surface adhesion coefficient calibration value. The lower the adhesion coefficient calibration value, the smaller the first deceleration slope value. The step size is determined based on the overshoot amount when the normalized game risk potential energy exceeds the first action threshold. The larger the overshoot amount, the larger the step size value. During braking, the normalized game risk potential energy is continuously monitored. When the normalized game risk potential energy is lower than the second action threshold and the duration exceeds the calibrated exit confirmation time, the braking device is controlled to stop increasing the braking pressure and gradually restore the deceleration to zero.

[0068] In this embodiment of the application, firstly, the historical risk data under the risk-free state is statistically analyzed to obtain the benchmark risk threshold: the normalized game risk potential energy historical values ​​of multiple consecutive time windows during risk-free following are collected, and the 90th percentile of these values ​​is calculated as the benchmark risk threshold. This value can cover the fluctuations of most risk-free scenarios and will not misjudge normal small fluctuations as risks.

[0069] Secondly, the speed increase coefficient is determined based on the current speed of the vehicle. The higher the current speed of the vehicle, the smaller the potential energy of the risk is enough to pose a collision threat. Therefore, the speed increase coefficient is set to be inversely proportional to the current speed of the vehicle. For example, the speed increase coefficient = calibrated reference speed / current speed of the vehicle, where the calibrated reference speed is 30 kilometers per hour.

[0070] Next, the baseline risk threshold is multiplied by the growth rate coefficient to obtain the final first action threshold, ensuring that braking is triggered earlier in high-speed scenarios and that braking is not triggered too frequently and sensitively in low-speed scenarios. Then, the ratio of the median of the historical value of the normalized game risk potential energy to the baseline risk threshold is calculated, and this ratio is multiplied by the first action threshold to obtain the second action threshold, thus completing the adaptive determination of the two action thresholds.

[0071] Furthermore, when the real-time normalized game risk potential energy exceeds the first action threshold, a graded braking operation is triggered. In the initial braking stage, braking is applied according to the first deceleration slope, which is determined by the current road surface adhesion coefficient calibration value. The lower the adhesion coefficient, the more slippery the road surface, and the smaller the initial slope should be set. For example, it can be calculated as: first deceleration slope = reference deceleration slope × (road surface adhesion coefficient calibration value / reference adhesion coefficient), where the reference adhesion coefficient is taken as the adhesion coefficient value of dry asphalt road surface (0.8), and the reference deceleration slope is taken as 3 m / s. 3 .

[0072] At fixed time intervals, the deceleration slope is increased by a step size, determined by the risk potential energy overshoot. For example, step size = baseline step size × (overshoot / baseline overshoot), where overshoot = normalized game risk potential energy - first action threshold. The baseline overshoot is set to 0.2, and the baseline step size is 0.5 m / s. 3 The larger the overshoot, the higher the risk. The larger the step size, the faster the deceleration increases, until the deceleration slope reaches the preset maximum value and then stops increasing.

[0073] During the braking process, the real-time risk potential energy is continuously monitored. When the normalized game risk potential energy is lower than the second action threshold and the duration of this state exceeds the preset exit confirmation time, the risk is confirmed to have been eliminated. The braking device is then controlled to stop increasing the braking pressure and gradually adjust the deceleration to zero to complete the braking exit operation, thus avoiding further deceleration that could affect the overall traffic efficiency.

[0074] Based on the above steps S60, the risk of cutting in can be dealt with in a timely manner by graded braking after the risk is triggered. At the same time, the braking strategy can be adaptively adjusted according to the risk changes, and the braking can be disengaged in a timely manner after the risk is eliminated. This balances driving safety and traffic efficiency, and realizes a closed loop of perception and decision-making for unstable road factors.

[0075] In summary, compared with existing technologies, this application quantifies and identifies the driver's intention to cut in line step by step, calculates the dynamic risk coefficient by combining the target vehicle's motion state, and further corrects the risk by combining the current road conditions and driving speed of the vehicle. Finally, it obtains an accurate quantitative result of road instability risk. This can effectively solve the technical defects of existing technologies that rely solely on vehicle motion trajectory parameters to predict cutting-in risk, do not consider the potential uncertainty brought about by the driver's subjective intention, and do not correct the risk assessment results by combining the actual driving conditions of the vehicle. It improves the perception accuracy of unstable factors such as cutting in line in autonomous driving scenarios, provides more reliable input for subsequent collision avoidance decisions, and matches an adaptive graded braking exit strategy, taking into account both driving safety and traffic efficiency.

[0076] In summary, the embodiments of this application have at least the following technical effects: This application provides an AI-based method for perceiving road instability factors in autonomous driving. First, by extracting the micro-motion characteristics of the target vehicle's steering wheel angle, it identifies premeditated cutting-in intentions before the act is completed, only during the driver's initial fine-tuning preparation stage. Second, by combining the target vehicle's instantaneous headway and lateral acceleration to obtain a dynamic risk coefficient, and then combining the current road surface adhesion conditions and the vehicle's current speed to correct the dynamic risk coefficient, a normalized game-theoretic risk potential energy matching the current driving scenario is obtained, adaptable to the risk response requirements of different road surfaces. Finally, a graded, incremental braking control strategy and an early exit mechanism are adopted to avoid the risk of vehicle skidding and loss of control caused by emergency braking, and also to avoid unnecessary continuous braking affecting traffic efficiency, effectively improving the driving safety of autonomous vehicles dealing with dynamic road instability factors on open roads.

[0077] Through the above technical solution, this application solves the technical problem that the existing technology cannot identify the driver's premeditated intention to cut in front of the vehicle before the cutting behavior is completed, and can only trigger the response after the vehicle enters the lane. This leaves insufficient reaction time for the autonomous vehicle and makes it difficult to cope with sudden risks under different road conditions. The new technology can perceive the dynamic instability factors of the road surface in advance, match different road surface adhesion conditions and vehicle speed to dynamically adjust the risk threshold and braking strategy, take into account driving safety and traffic efficiency, and adapt to the driving needs of various complex road surface scenarios in open roads.

[0078] Example 2, as Figure 3 As shown, based on the same inventive concept as the AI-based autonomous driving road instability factor perception method provided in Embodiment 1, this application also provides an AI-based autonomous driving road instability factor perception system, including: The vehicle parameter acquisition module 11 is used to acquire in real time the steering wheel angle micro-motion sequence of target vehicles around the autonomous vehicle, the instantaneous head-to-head distance between the front of the autonomous vehicle and the rear bumper of the target vehicle, and the lateral slip speed of the target vehicle relative to the center line of the lane. The steering wheel angle extraction module 12 is used to extract the steering wheel angle micro-motion sequence within the previous time window when the lateral slip speed is detected to be greater than zero and the instantaneous headway is continuously reduced beyond the dynamic headway compression rate threshold. The lane-cutting intention prediction module 13 is used to calculate the proportion of the zero speed range of the steering wheel and the frequency of the steering wheel sign reversal based on the steering wheel angle micro-motion sequence, and input them into the lane-cutting intention prediction model to output the confidence level of the driver's intention to cut in line. The dynamic risk coefficient acquisition module 14 is used to acquire a dynamic risk coefficient representing the urgency of intrusion based on the intent confidence and the instantaneous headway of the vehicle, and in combination with the lateral acceleration of the target vehicle. The dynamic risk coefficient correction module 15 is used to attenuate and correct the dynamic risk coefficient based on the adhesion coefficient of the road surface where the unmanned vehicle is currently located and the current vehicle speed, so as to obtain the normalized game risk potential energy. The graded braking execution module 16 is used to control the autonomous vehicle to perform graded braking operation with deceleration increasing over time when the normalized game risk potential energy exceeds the first action threshold, and to exit the graded braking operation in advance when the normalized game risk potential energy falls back to below the second action threshold.

[0079] In one embodiment, the vehicle parameter acquisition module 11 is specifically used for: The vehicle continuously acquires images of the front tire treads of the target vehicle using an onboard camera, and extracts estimated values ​​of the front wheel steering angle from the images of the front tire treads. The estimated value of the front wheel steering angle is recorded at a fixed sampling period, and combined with the steering system angular transmission ratio to generate the instantaneous value of the steering wheel angle. The steering system angular transmission ratio is the ratio of the maximum rotation angle of the target vehicle's steering wheel to the maximum mechanical deflection angle of the front wheels. The instantaneous values ​​of steering wheel angle obtained from N consecutive sampling periods are arranged in chronological order to form the steering wheel angle micro-motion sequence, where N is an integer between 50 and 100; The longitudinal relative distance between the front of the autonomous vehicle and the rear bumper of the target vehicle is measured synchronously by millimeter-wave radar, the current speed of the autonomous vehicle is read, and the instantaneous front-to-rear distance is calculated. The vehicle-mounted camera identifies the position of the lane centerline, detects the lateral offset of the midpoint of the target vehicle in the direction perpendicular to the lane centerline, and divides the difference of the lateral offset obtained in two consecutive sampling periods by the sampling time interval to obtain the lateral slip velocity of the target vehicle relative to the lane centerline. The instantaneous values ​​of steering wheel angle, instantaneous headway, and lateral slip speed collected at the same sampling moment are aligned by timestamps to form an input parameter set synchronized with the time window of the steering wheel angle micro-motion sequence.

[0080] Furthermore, by continuously acquiring images of the front tire treads of the target vehicle using an onboard camera, and extracting estimated values ​​of the front wheel steering angle from the front tire tread images, including: Identify the boundary line between the wheel hub edge and the tire tread contact area from the front wheel tire tread image, extract at least two feature points at different radial positions on the boundary line, and connect the feature points to form a wheel hub reference line; Using the horizontal axis of the front tire tread image as the reference direction of the vehicle's longitudinal axis, calculate the angle between the wheel hub reference line and the horizontal axis of the image, and use the angle as the initial front wheel steering angle; Extract the hub center point and the tire tread pattern symmetry axis from the front tire tread image, and calculate the offset angle of the symmetry axis relative to the hub center point and the direction perpendicular to the horizontal axis of the image, as the steering correction angle; The estimated value of the front wheel steering angle is obtained by subtracting the steering correction angle from the initial front wheel steering angle.

[0081] Furthermore, in one embodiment of the application, the calculation of the dynamic time-distance compression ratio threshold includes: Continuous sampling points whose absolute values ​​of the instantaneous steering wheel angle are less than the steering angle silence threshold are extracted from the steering wheel angle micro-motion sequence to form a silence segment. The steering angle silence threshold is calibrated based on the steering wheel angle sensor noise level and the steering device free travel. The instantaneous headway distance of each sampling point within the silent segment is statistically analyzed, and the average attenuation slope of the instantaneous headway distance within the silent segment is calculated. Combined with the time distance compression amplification factor, the basic compression ratio threshold is calculated. The calculation steps for the time-distance compression amplification factor include: The attenuation difference is obtained by subtracting the instantaneous headway from the end of the sampling point from the instantaneous headway at the beginning of the silent segment. The relative attenuation amplitude is determined based on the attenuation difference and the instantaneous headway at the starting sampling point; Add 1 to the relative attenuation amplitude and use it as the time-distance compression amplification factor; The change in instantaneous value of steering wheel angle between two adjacent sampling points is extracted from the steering wheel angle micro-motion sequence, and the rate of change of angle is calculated by combining the sampling time interval. The duration of all corner change rates exceeding the rate threshold within the time window is summed, and the proportion of aggressive operations is calculated by combining the total duration of the time window. The rate threshold is a calibrated rate threshold that distinguishes between aggressive and normal operations based on historical driving data statistics. Collect data from a large-scale natural driving database, including the rate of change of steering wheel angle of a large number of drivers in a stable following state, and calculate the proportion of aggressive operations in each time window; The percentages of aggressive operations were sorted from smallest to largest, and the 85th percentile was taken as the upper limit of the statistics. The dynamic time-distance compression ratio threshold is determined based on the base compression ratio threshold, the proportion of aggressive operations, and the statistical upper limit value, which are calculated continuously in real time. The base compression ratio threshold and the proportion of aggressive operations are updated online according to the steering wheel angle micro-motion sequence maintained by continuous sliding window. The dynamic time-distance compression ratio threshold is directly proportional to the base compression ratio threshold and the proportion of aggressive operations, and inversely proportional to the statistical upper limit value.

[0082] Further, in one embodiment of the application, based on the steering wheel angle micro-motion sequence, calculating the proportion of the steering wheel zero-speed range and the steering angle sign reversal frequency includes: Traverse the steering wheel angle micro-motion sequence, compare the instantaneous value of the steering wheel angle at each sampling point with the angle silence threshold, and mark the zero speed point where the absolute value of the angle is less than the angle silence threshold; Divide consecutive adjacent zero-velocity points within the time window into several zero-velocity segments, and sum the sampling duration spanned by each zero-velocity segment to obtain the total zero-velocity duration. Calculate the ratio of the total zero-speed time to the total time of the time window, and use it as the proportion of the zero-speed range of the steering wheel; Based on the steering wheel angle micro-motion sequence, the sampling points marked as zero speed points are removed to obtain the non-zero speed point sequence; Iterate through the instantaneous steering wheel angle values ​​of two adjacent sampling points in the non-zero speed point sequence, and check whether the signs of the angle values ​​of the two adjacent points are opposite. If the signs are opposite, count one angle sign reversal event. The total number of corner sign reversal events occurring within the time window is accumulated, and the ratio of the total number to the total number of adjacent sampling point pairs in the non-zero velocity point sequence is calculated as the corner sign reversal frequency.

[0083] Furthermore, the construction steps of the preemptive cut-in intention prediction model include: Collect multiple instances of target vehicles cutting in front of your vehicle from adjacent lanes during historical driving, as well as an equal number of normal following events; The proportion of the steering wheel's zero-speed range and the frequency of the turning sign reversal within the time window preceding each event are extracted to form a sample feature set. Each sample feature set is labeled with a corresponding label for the actual lane-cutting intention. The label for lane-cutting events is set to 1, and the label for normal following events is set to 0. Based on the logistic regression algorithm, an initial lane-cutting intention prediction model is constructed. The initial lane-cutting intention prediction model uses the proportion of the steering wheel zero speed range and the frequency of the steering angle sign reversal as two input features. The two features are multiplied by their corresponding weights and then added together. The sum is then mapped to the 0 to 1 range by the Sigmoid function, and the output value is used as the confidence level of the driver's intention to cut in. The proportion of the steering wheel zero-speed interval and the frequency of the corner sign reversal in the sample feature set are input into the initial lane-cutting intention prediction model to obtain the prediction intention confidence. The binary cross-entropy loss between the prediction intention confidence and the actual lane-cutting intention label is calculated. With minimizing the binary cross-entropy loss as the training objective, the gradient descent method is used to iteratively optimize the two weight parameters in the initial preemptive intention prediction model. After each iteration, the average cross-entropy loss on the validation set is calculated. Training stops when the average value of the cross-entropy loss decreases by less than the preset convergence threshold for several consecutive iterations, and the current weight parameters are fixed to obtain the trained model for predicting the intention to cut in line.

[0084] In one embodiment, the dynamic risk coefficient acquisition module 14 is specifically used for: The reciprocal of the instantaneous headway is taken as the longitudinal approach rate, and combined with the intention confidence, a weighted longitudinal approximation value is determined. The lateral acceleration of the target vehicle is obtained, and the absolute value of the lateral acceleration is taken as the lateral disturbance intensity. The lateral acceleration is determined based on the ratio of the change in the lateral slip velocity at adjacent sampling times to the sampling time interval. Based on the intent confidence level, horizontal and vertical weights are determined, wherein the horizontal weight decreases as the intent confidence level increases, and the vertical weight increases as the intent confidence level increases. Based on the horizontal weights and the vertical weights, the weighted vertical approximation value and the horizontal interference intensity are weighted and fused to obtain the dynamic risk coefficient.

[0085] In one embodiment, the dynamic risk coefficient correction module 15 is specifically used for: Read the adhesion coefficient calibration value of the road surface where the unmanned vehicle is currently located; Read the current speed of the autonomous vehicle and use it as the current speed of the driverless vehicle; Based on the adhesion coefficient calibration value and the current vehicle speed, a vehicle speed-road surface joint factor is determined, wherein the vehicle speed-road surface joint factor is inversely proportional to the adhesion coefficient calibration value and directly proportional to the current vehicle speed; Multiplying the dynamic risk coefficient by the vehicle speed-road surface joint factor yields the normalized game risk potential energy.

[0086] In one embodiment, the graded braking execution module 16 is specifically used for: Collect the historical normalized game risk potential energy of the unmanned vehicle in a risk-free following state for multiple consecutive time windows, and calculate the 90th percentile of the historical normalized game risk potential energy as the benchmark risk threshold. Based on the current speed of the vehicle, an acceleration coefficient is obtained, wherein the acceleration coefficient is inversely proportional to the current speed of the vehicle; Based on the growth rate coefficient and the benchmark risk threshold, a first action threshold is obtained; The ratio of the median of the historical normalized game risk potential energy to the benchmark risk threshold is calculated and used as the exit ratio coefficient. Combined with the first action threshold, the second action threshold is determined. When the normalized game risk potential energy calculated in real time exceeds the first action threshold, graded braking is triggered. During the initial braking phase, braking is applied according to the first deceleration slope. After each fixed time interval, the currently used deceleration slope is increased by one step until the maximum deceleration slope is reached. The first deceleration slope is determined based on the current road surface adhesion coefficient calibration value. The lower the adhesion coefficient calibration value, the smaller the first deceleration slope value. The step size is determined based on the overshoot amount when the normalized game risk potential energy exceeds the first action threshold. The larger the overshoot amount, the larger the step size value. During braking, the normalized game risk potential energy is continuously monitored. When the normalized game risk potential energy is lower than the second action threshold and the duration exceeds the calibrated exit confirmation time, the braking device is controlled to stop increasing the braking pressure and gradually restore the deceleration to zero.

Claims

1. An artificial intelligence-based method for perceiving road instability factors in autonomous driving, characterized in that, The method includes: The system acquires in real time the steering wheel angle micro-motion sequence of target vehicles around the autonomous vehicle, the instantaneous head-to-head distance between the autonomous vehicle's front and the target vehicle's rear bumper, and the lateral slip speed of the target vehicle relative to the lane centerline. When the lateral slip velocity is detected to be greater than zero and the instantaneous headway is continuously reduced beyond the dynamic headway compression rate threshold, the steering wheel angle micro-motion sequence within the previous time window at the current moment is extracted. Based on the steering wheel angle micro-motion sequence, the proportion of the steering wheel zero speed range and the frequency of the steering wheel sign reversal are calculated, and then input into the lane-cutting intention prediction model to output the confidence level of the driver's intention to cut in. Based on the intent confidence and the instantaneous headway, and combined with the lateral acceleration of the target vehicle, a dynamic risk coefficient characterizing the urgency of the intrusion is obtained. Based on the adhesion coefficient of the road surface where the autonomous vehicle is currently located and the current speed of the vehicle, the dynamic risk coefficient is attenuated and corrected to obtain the normalized game risk potential energy. When the normalized game risk potential energy exceeds the first action threshold, the driverless vehicle is controlled to perform a graded braking operation with deceleration increasing over time. When the normalized game risk potential energy falls back below the second action threshold, the graded braking operation is prematurely terminated.

2. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 1, characterized in that, Real-time acquisition of the steering wheel angle micro-motion sequence of target vehicles surrounding the autonomous vehicle, the instantaneous headway between the autonomous vehicle's front and the target vehicle's rear bumper, and the lateral slip velocity of the target vehicle relative to the lane centerline, including: The vehicle continuously acquires images of the front tire treads of the target vehicle using an onboard camera, and extracts estimated values ​​of the front wheel steering angle from the images of the front tire treads. The estimated value of the front wheel steering angle is recorded at a fixed sampling period, and combined with the steering system angular transmission ratio to generate the instantaneous value of the steering wheel angle. The steering system angular transmission ratio is the ratio of the maximum rotation angle of the target vehicle's steering wheel to the maximum mechanical deflection angle of the front wheels. The instantaneous values ​​of steering wheel angle obtained from N consecutive sampling periods are arranged in chronological order to form the steering wheel angle micro-motion sequence, where N is an integer between 50 and 100; The longitudinal relative distance between the front of the autonomous vehicle and the rear bumper of the target vehicle is measured synchronously by millimeter-wave radar, the current speed of the autonomous vehicle is read, and the instantaneous front-to-rear distance is calculated. The vehicle-mounted camera identifies the position of the lane centerline, detects the lateral offset of the midpoint of the target vehicle in the direction perpendicular to the lane centerline, and divides the difference of the lateral offset obtained in two consecutive sampling periods by the sampling time interval to obtain the lateral slip velocity of the target vehicle relative to the lane centerline. The instantaneous values ​​of steering wheel angle, instantaneous headway, and lateral slip speed collected at the same sampling moment are aligned by timestamps to form an input parameter set synchronized with the time window of the steering wheel angle micro-motion sequence.

3. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 2, characterized in that, The vehicle continuously acquires images of the front tire treads of the target vehicle using an onboard camera, and extracts estimated front wheel steering angle values ​​from the front tire tread images, including: Identify the boundary line between the wheel hub edge and the tire tread contact area from the front wheel tire tread image, extract at least two feature points at different radial positions on the boundary line, and connect the feature points to form a wheel hub reference line; Using the horizontal axis of the front tire tread image as the reference direction of the vehicle's longitudinal axis, calculate the angle between the wheel hub reference line and the horizontal axis of the image, and use the angle as the initial front wheel steering angle; Extract the hub center point and the tire tread pattern symmetry axis from the front tire tread image, and calculate the offset angle of the symmetry axis relative to the hub center point and the direction perpendicular to the horizontal axis of the image, as the steering correction angle; The estimated value of the front wheel steering angle is obtained by subtracting the steering correction angle from the initial front wheel steering angle.

4. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 1, characterized in that, The calculation of the dynamic time-distance compression ratio threshold includes: Continuous sampling points whose absolute values ​​of the instantaneous steering wheel angle are less than the steering angle silence threshold are extracted from the steering wheel angle micro-motion sequence to form a silence segment. The steering angle silence threshold is calibrated based on the steering wheel angle sensor noise level and the steering device free travel. The instantaneous headway distance of each sampling point within the silent segment is statistically analyzed, and the average attenuation slope of the instantaneous headway distance within the silent segment is calculated. Combined with the time distance compression amplification factor, the basic compression ratio threshold is calculated. The calculation steps for the time-distance compression amplification factor include: The attenuation difference is obtained by subtracting the instantaneous headway from the end of the sampling point from the instantaneous headway at the beginning of the silent segment. The relative attenuation amplitude is determined based on the attenuation difference and the instantaneous headway at the starting sampling point; Add 1 to the relative attenuation amplitude and use it as the time-distance compression amplification factor; The change in instantaneous value of steering wheel angle between two adjacent sampling points is extracted from the steering wheel angle micro-motion sequence, and the rate of change of angle is calculated by combining the sampling time interval. The duration of all corner change rates exceeding the rate threshold within the time window is summed, and the proportion of aggressive operations is calculated by combining the total duration of the time window. The rate threshold is a calibrated rate threshold that distinguishes between aggressive and normal operations based on historical driving data statistics. Collect data from a large-scale natural driving database, including the rate of change of steering wheel angle of a large number of drivers in a stable following state, and calculate the proportion of aggressive operations in each time window; The percentages of aggressive operations were sorted from smallest to largest, and the 85th percentile was taken as the upper limit of the statistics. The dynamic time-distance compression ratio threshold is determined based on the base compression ratio threshold, the proportion of aggressive operations, and the statistical upper limit value, which are calculated continuously in real time. The base compression ratio threshold and the proportion of aggressive operations are updated online according to the steering wheel angle micro-motion sequence maintained by continuous sliding window. The dynamic time-distance compression ratio threshold is directly proportional to the base compression ratio threshold and the proportion of aggressive operations, and inversely proportional to the statistical upper limit value.

5. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 1, characterized in that, Based on the aforementioned steering wheel angle micro-motion sequence, the percentage of the steering wheel's zero-speed range and the frequency of angle sign reversal are calculated, including: Traverse the steering wheel angle micro-motion sequence, compare the instantaneous value of the steering wheel angle at each sampling point with the angle silence threshold, and mark the zero speed point where the absolute value of the angle is less than the angle silence threshold; Divide consecutive adjacent zero-velocity points within the time window into several zero-velocity segments, and sum the sampling duration spanned by each zero-velocity segment to obtain the total zero-velocity duration. Calculate the ratio of the total zero-speed time to the total time of the time window, and use it as the proportion of the zero-speed range of the steering wheel; Based on the steering wheel angle micro-motion sequence, the sampling points marked as zero speed points are removed to obtain the non-zero speed point sequence; Iterate through the instantaneous steering wheel angle values ​​of two adjacent sampling points in the non-zero speed point sequence, and check whether the signs of the angle values ​​of the two adjacent points are opposite. If the signs are opposite, count one angle sign reversal event. The total number of corner sign reversal events occurring within the time window is accumulated, and the ratio of the total number to the total number of adjacent sampling point pairs in the non-zero velocity point sequence is calculated as the corner sign reversal frequency.

6. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 1, characterized in that, The steps for constructing the model for predicting the intention to cut in line include: Collect multiple instances of target vehicles cutting in front of your vehicle from adjacent lanes during historical driving, as well as an equal number of normal following events; The proportion of the steering wheel's zero-speed range and the frequency of the turning sign reversal within the time window preceding each event are extracted to form a sample feature set. Each sample feature set is labeled with a corresponding label for the actual lane-cutting intention. The label for lane-cutting events is set to 1, and the label for normal following events is set to 0. Based on the logistic regression algorithm, an initial lane-cutting intention prediction model is constructed. The initial lane-cutting intention prediction model uses the proportion of the steering wheel zero speed range and the frequency of the steering angle sign reversal as two input features. The two features are multiplied by their corresponding weights and then added together. The sum is then mapped to the 0 to 1 range by the Sigmoid function, and the output value is used as the confidence level of the driver's intention to cut in. The proportion of the steering wheel zero-speed interval and the frequency of the corner sign reversal in the sample feature set are input into the initial lane-cutting intention prediction model to obtain the prediction intention confidence. The binary cross-entropy loss between the prediction intention confidence and the actual lane-cutting intention label is calculated. With minimizing the binary cross-entropy loss as the training objective, the gradient descent method is used to iteratively optimize the two weight parameters in the initial preemptive intention prediction model. After each iteration, the average cross-entropy loss on the validation set is calculated. Training stops when the average value of the cross-entropy loss decreases by less than the preset convergence threshold for several consecutive iterations, and the current weight parameters are fixed to obtain the trained model for predicting the intention to cut in line.

7. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 1, characterized in that, Based on the intent confidence level and the instantaneous headway, and combined with the lateral acceleration of the target vehicle, a dynamic risk coefficient characterizing the urgency of the intrusion is obtained, including: The reciprocal of the instantaneous headway is taken as the longitudinal approach rate, and combined with the intention confidence, a weighted longitudinal approximation value is determined. The lateral acceleration of the target vehicle is obtained, and the absolute value of the lateral acceleration is taken as the lateral disturbance intensity. The lateral acceleration is determined based on the ratio of the change in the lateral slip velocity at adjacent sampling times to the sampling time interval. Based on the intent confidence level, horizontal and vertical weights are determined, wherein the horizontal weight decreases as the intent confidence level increases, and the vertical weight increases as the intent confidence level increases. Based on the horizontal weights and the vertical weights, the weighted vertical approximation value and the horizontal interference intensity are weighted and fused to obtain the dynamic risk coefficient.

8. The method for perceiving road instability factors in autonomous driving based on artificial intelligence according to claim 1, characterized in that, Based on the adhesion coefficient of the road surface where the autonomous vehicle is currently located and the vehicle's current speed, the dynamic risk coefficient is attenuated and corrected to obtain the normalized game risk potential energy, including: Read the adhesion coefficient calibration value of the road surface where the unmanned vehicle is currently located; Read the current speed of the autonomous vehicle and use it as the current speed of the driverless vehicle; Based on the adhesion coefficient calibration value and the current vehicle speed, a vehicle speed-road surface joint factor is determined, wherein the vehicle speed-road surface joint factor is inversely proportional to the adhesion coefficient calibration value and directly proportional to the current vehicle speed; Multiplying the dynamic risk coefficient by the vehicle speed-road surface joint factor yields the normalized game risk potential energy.

9. The method for perceiving road instability factors based on artificial intelligence for unmanned driving according to claim 1, characterized in that, When the normalized game risk potential energy exceeds the first action threshold, the autonomous vehicle is controlled to perform a graded braking operation with deceleration increasing over time. When the normalized game risk potential energy falls below the second action threshold, the graded braking operation is prematurely terminated, including: Collect the historical normalized game risk potential energy of the unmanned vehicle in a risk-free following state for multiple consecutive time windows, and calculate the 90th percentile of the historical normalized game risk potential energy as the benchmark risk threshold. Based on the current speed of the vehicle, an acceleration coefficient is obtained, wherein the acceleration coefficient is inversely proportional to the current speed of the vehicle; Based on the growth rate coefficient and the benchmark risk threshold, a first action threshold is obtained; The ratio of the median of the historical normalized game risk potential energy to the benchmark risk threshold is calculated and used as the exit ratio coefficient. Combined with the first action threshold, the second action threshold is determined. When the normalized game risk potential energy calculated in real time exceeds the first action threshold, graded braking is triggered. During the initial braking phase, braking is applied according to the first deceleration slope. After each fixed time interval, the currently used deceleration slope is increased by one step until the maximum deceleration slope is reached. The first deceleration slope is determined based on the current road surface adhesion coefficient calibration value. The lower the adhesion coefficient calibration value, the smaller the first deceleration slope value. The step size is determined based on the overshoot amount when the normalized game risk potential energy exceeds the first action threshold. The larger the overshoot amount, the larger the step size value. During braking, the normalized game risk potential energy is continuously monitored. When the normalized game risk potential energy is lower than the second action threshold and the duration exceeds the calibrated exit confirmation time, the braking device is controlled to stop increasing the braking pressure and gradually restore the deceleration to zero.

10. An artificial intelligence-based autonomous driving road instability factor perception system, characterized in that, The method for performing the artificial intelligence-based autonomous driving road instability factor perception method according to any one of claims 1-9 includes: The vehicle parameter acquisition module is used to acquire in real time the steering wheel angle micro-motion sequence of target vehicles around the autonomous vehicle, the instantaneous head-to-head distance between the front of the autonomous vehicle and the rear bumper of the target vehicle, and the lateral slip speed of the target vehicle relative to the center line of the lane. The steering wheel angle extraction module is used to extract the steering wheel angle micro-motion sequence within the previous time window when the lateral slip speed is detected to be greater than zero and the instantaneous headway is continuously reduced beyond the dynamic headway compression rate threshold. The lane-cutting intention prediction module is used to calculate the proportion of the zero-speed range of the steering wheel and the reversal frequency of the steering wheel sign based on the steering wheel angle micro-motion sequence, and input them into the lane-cutting intention prediction model to output the confidence level of the driver's intention to cut in. The dynamic risk coefficient acquisition module is used to acquire a dynamic risk coefficient representing the urgency of intrusion based on the intent confidence and the instantaneous headway of the vehicle, and in combination with the lateral acceleration of the target vehicle. The dynamic risk coefficient correction module is used to attenuate and correct the dynamic risk coefficient based on the adhesion coefficient of the road surface where the unmanned vehicle is currently located and the current speed of the vehicle, so as to obtain the normalized game risk potential energy. The graded braking execution module is used to control the autonomous vehicle to perform graded braking operations with increasing deceleration over time when the normalized game risk potential energy exceeds the first action threshold, and to exit the graded braking operation in advance when the normalized game risk potential energy falls back below the second action threshold.

Images