Method for predicting takeover success probability and evaluating intervention effect under human-machine co-driving condition

Abstract

The application relates to the field of automatic driving, and particularly discloses a method for predicting takeover success probability and evaluating intervention effect under human-machine co-driving conditions. The method comprises the following steps: step 1, establishing a takeover success sample: determining the optimal takeover budget time and the takeover success rate cumulative curve of a driver, and constructing a mapping parameter set; step 2, predicting the takeover fitness of the driver: optimizing parameters and constructing a real-time takeover state vector T, calculating the Mahalanobis distance D M (S) of the real-time takeover state vector and a perfect takeover state vector D M (F), calculating the basic trust probability assignment m(s) of perfect takeover and the basic trust probability assignment m(f) of a failed takeover, and setting the decision threshold of the trust degree assignment as epsilon, and predicting that the driver can successfully take over when m(s)-m(f) > epsilon. The method for predicting takeover success probability and evaluating intervention effect under human-machine co-driving conditions can improve prediction accuracy and efficiency, and reduce takeover risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic driving, in particular to a method for takeover success probability prediction and intervention effect evaluation under human-machine co-driving condition. BACKGROUND

[0002] For vehicles with L3 or L4 level automatic driving function, when in the automatic driving stage, the human driver needs to take over the control of the vehicle in time due to the limitations or failures of the automatic driving system. If the driver does not correctly place his hands and feet on the steering wheel or brake pedal or does not observe the surrounding environment within a certain time, and is not familiar with the motion state of the vehicle and the driving environment, the driver will recklessly transfer the control of the vehicle to the human driver, which may cause accidents due to judgment and decision-making errors or failure to avoid obstacles in front of the vehicle. Therefore, the probability of successful takeover by the driver needs to be predicted before the control is transferred, and when the probability of successful takeover is low, the automatic driving vehicle should take risk mitigation measures in time to avoid accidents caused by failed takeover and improve the safety of manual takeover.

[0003] The current common judgment method, such as the patent with application number CN202111116691.8: a method for predicting the takeover ability of a driver under an automatic driving system, generally tests the driving ability, reaction speed, situational awareness, fatigue index, and emotional state of the driver under different takeover abilities in advance and enters them into a database, and obtains the ROC curve of each driving feature index of different drivers. In prediction, the current state of the driver is combined with the data in the database for calculation, and finally the score of the driver in terms of actual driving ability, reaction speed, etc. is determined, and the corresponding rating standard is referred to. If the rating is low, the driver is refused to take over.

[0004] The above prediction method can detect some dangerous takeover behaviors of the driver in advance, but it only refuses the driver to take over according to the rating once, and when the driver still needs to take over the control, the driver needs to make another takeover request, which leads to low efficiency and comfort of takeover. SUMMARY

[0005] The present application aims to provide a method for takeover success probability prediction and intervention effect evaluation under human-machine co-driving condition to improve the comfort of takeover and reduce the risk of takeover.

[0006] To achieve the above purpose, the present application adopts the following technical solution: a method for takeover success probability prediction and intervention effect evaluation under human-machine co-driving condition, comprising the following steps:

[0007] Step 1, establishing a takeover success sample: determining the optimal takeover budget time and takeover success rate cumulative curve of the driver in any takeover scene through pre-test, and constructing a mapping parameter set in the takeover scene for different takeover behaviors and facing the whole takeover process;

[0008] Step 2, predicting the takeover fitness of the driver: dynamically collecting the head-eye movement, body position, psychological load, vehicle motion parameters and driving environment state parameters of the driver, optimizing the parameters through mayfly algorithm and constructing a real-time takeover state vector T:

[0009]

[0010] The Mahalanobis distance D between the real-time takeover state vector and the perfect takeover state vector is calculated by formula (1) M (S):

[0011]

[0012] The Mahalanobis distance D between the real-time takeover state vector and the failure takeover state vector is calculated by formula (2) M (F):

[0013]

[0014] The basic trust probability assignment m(s) of perfect takeover and the basic trust probability assignment m(f) of failure takeover are calculated by formula (3)

[0015]

[0016] And the decision threshold of the trust degree assignment is set as ε, when m(s)-m(f) > ε, it is predicted that the driver can successfully take over.

[0017] The beneficial effects of the scheme are:

[0018] In the scheme, a represents the dimension of the scene takeover performance mapping parameter, R represents the characteristic parameter, μ represents the mean of the real-time state vector, and P represents the corresponding mapping parameter. In the scheme, the takeover sample is first established for the driver. In the test, whether the driver successfully takes over enables the takeover to be more accurately predicted according to the driving level of the driver in the prediction, effectively improving the accuracy of the prediction. In the scheme, the decision threshold ε of the trust degree assignment is set when predicting the takeover success rate. In the actual prediction, the decision threshold ε can be determined according to the takeover test of a number of drivers, or different decision thresholds ε can be set for drivers with different driving levels after the drivers are classified, so that the prediction method of the scheme can adapt to different drivers, further ensuring high prediction accuracy and reducing takeover risk.

[0019] The optimal takeover budget time in the scheme refers to the time at which the takeover signal is sent out to the driver until the driver starts to take over, and the takeover success rate is the highest; and the takeover success rate cumulative curve refers to a curve between the takeover success rate of the driver and the takeover time since the information is published.

[0020] Further, step 2 sends the takeover signal to the driver when m(s)-m(f)≤ε, and slides to the next time window to repeat the operation of step 2 for continuous prediction, and predicts that the driver can successfully take over when m(s)-m(f)>ε.

[0021] The beneficial effect of the scheme is that when the driver proposes a takeover demand, the scheme can predict again in the next time window even if the first prediction fails, without the driver repeating the operation of proposing the takeover demand, so that the operation of the driver is more convenient, the takeover efficiency is improved, and the safety of driving is improved by avoiding ignoring the surrounding road conditions due to the driver repeatedly proposing the operation of the takeover demand. The next time window refers to the time interval between adjacent two predictions based on one takeover request of the driver.

[0022] Further, step 1 determines the limit takeover time for completing the takeover during the test; and step 2, when continuously predicting, predicts that the driver fails to take over when m(s)-m(f)≤ε is still satisfied when sliding to the limit takeover time.

[0023] The beneficial effect of the scheme is that the limit takeover time refers to the time from when the takeover signal is sent out to the driver until the driver will be unable to complete the takeover; and reaching the limit takeover time indicates that the driver will not have enough time to take over, so even if the driving right is transferred to the driver, the driver is easy to cause danger due to untimely reaction, and the prediction method of the scheme determines that the takeover fails at this time, which effectively avoids such risks and improves the safety of driving.

[0024] Further, each takeover result is included in the takeover sample of step 1.

[0025] The beneficial effect of the scheme is that after each prediction result is included in the takeover sample, the driving level of the driver can be updated in real time, and after the level of the driver is improved, higher prediction accuracy is ensured.

[0026] Further, the method further comprises step 3 of intervening and regulating the takeover performance, and when m(s)-m(f)≤ε in step 4, at least one of the following is corrected: reinforcement training, takeover monitoring, auxiliary information, and signal optimization, wherein the reinforcement training comprises strengthening at least one of the visual search, mental load, reaction time, motion stability, and accuracy of the driver in the form of repeated practice; the takeover monitoring comprises reminding the driver to observe the traffic environment during the process from the application of the driver to transfer the takeover control to the formal transfer of the takeover control; the auxiliary information comprises marking the road conditions around the vehicle through a screen; and the signal optimization comprises prompting the driver in at least one of the following ways: sound, light, and vibration.

[0027] The beneficial effects of the present scheme are that the intervention and regulation of the takeover performance can be used to train the deficiencies of the driver, improve the driving level of the driver, and finally improve the safety of the takeover.

[0028] Further, the method further comprises step 4 of determining the scene complexity, wherein the operation mode of any one scene in step 1 is divided into only braking, only steering, braking first and then steering, steering first and then braking, and braking and steering simultaneously, a training sample set is established, the index complexity is normalized, the difficulty representation coefficient of each operation mode is obtained in the interval of 0 to 1, and the weight coefficient d of the scene is determined according to the difficulty representation coefficient.

[0029] Step 5 is to evaluate the takeover performance of the driver, wherein the takeover performance E1 of the specific takeover behavior of the driver is evaluated based on the test in step 1 by using formula (4).

[0030]

[0031] wherein Q is the characteristic value of each parameter of the driver in the scene; and after at least one intervention in step 3, the takeover performance E2 after the intervention is calculated by using the above formula.

[0032] The beneficial effects of the present scheme are that after the intervention and regulation of the takeover performance, the takeover performance after the regulation can be calculated by using steps 4 and 5, and when E2 is greater than E1, it is proved that the regulation method is suitable for the driver, and the regulation method can be continuously used to continuously improve the takeover success rate of the driver.

[0033] Further, step 3 generates a targeted regulation matrix M of the driver in any one takeover scene according to the test in step 1.

[0034]

[0035] And after the correction, the updated target control matrix M' of the driver in any one takeover scene is generated according to the test of step 1:

[0036]

[0037] And the weight vector R of the to-be-controlled parameter of any one takeover scene is determined according to the distance between the best and worst solutions, and R=(r1, r2,..., r n ) T , and the variation coefficient of a single parameter is calculated as C kn using formula (5):

[0038]

[0039] And when C kn is a positive index and increases after correction, C kn takes a positive value; and decreases after correction, C kn takes a negative value; and when C kn is a negative index and decreases after correction, C kn takes a positive value; and increases after correction, C kn takes a negative value; the variation coefficient matrix C is obtained to represent the improvement effect of any one takeover scene:

[0040]

[0041] The improvement effect Y of any one takeover scene efficiency is represented by formula (6):

[0042] Y=C·R (6)

[0043] When Y≥N, the correction method is used for correction of all scenes.

[0044] The beneficial effects of the present scheme are: k is the number of takeover scenes, n is the number of to-be-controlled parameters, and J is the value of a takeover performance representation parameter in a specified takeover scene; wherein N is a constant and is set by a person. The method of the present scheme can accurately find out the shortcomings of the driver and effectively improve the pertinence of training.

[0045] Further, step 6, when determining the takeover performance, performs mean value processing on all takeover performances of the driver in the same scene to obtain a mean value takeover performance evaluation result E1', and E1' is taken as the takeover performance corresponding to the specific takeover behavior of the driver in step 3.

[0046] The beneficial effects of the present scheme are: mean value processing can reduce the error of the test and further improve the accuracy of the prediction. DETAILED DESCRIPTION

[0047] The following is further described in detail through a specific embodiment:

[0048] Embodiment 1

[0049] The method for takeover success probability prediction and intervention effect evaluation under human-machine co-driving conditions comprises the following steps:

[0050] Step 1, establishing a takeover success sample: determining the optimal takeover budget time, takeover success rate cumulative curve and limit takeover time of the driver in any takeover scenario through pre-test, and constructing a mapping parameter set under each scenario for different takeover behaviors and facing the entire takeover process;

[0051] Step 2, driver takeover fitness prediction: dynamically collecting driver's head-eye movement, body position, psychological load, vehicle motion parameters and driving environment state parameters, optimizing parameters through mayfly algorithm and constructing a real-time takeover state vector T:

[0052]

[0053] The Mahalanobis distance D between the real-time takeover state vector and the perfect takeover state vector is calculated by formula (1) M (S):

[0054]

[0055] The Mahalanobis distance D between the real-time takeover state vector and the failure takeover state vector is calculated by formula (2) M (F):

[0056]

[0057] The basic trust probability assignment m(s) of perfect takeover and the basic trust probability assignment m(f) of failure takeover are calculated by formula (3):

[0058]

[0059] And set the decision threshold of trust degree assignment as ε, when m(s)-m(f) > ε, it is predicted that the driver can take over successfully; when m(s)-m(f) ≤ ε, send a takeover signal to the driver, and slide to the next time window to repeat the operation of step 2 for continuous prediction; and when m(s)-m(f) > ε, it is predicted that the driver can take over successfully; when sliding to the limit takeover time still satisfies m(s)-m(f) ≤ ε, it is predicted that the driver fails to take over; every time a takeover is completed, the takeover result is included in the takeover sample of step 1. Each time window in this embodiment is 0.1s, and ε is 0.4-0.5, and ε in this embodiment is 0.5.

[0060] Step 3, intervention and regulation of takeover performance: when m(s)-m(f)≤ε in step 2, correction is made from at least one of the dimensions of reinforcement training, monitoring takeover, auxiliary information and signal optimization, the reinforcement training includes strengthening at least one of the visual search, mental load, reaction time, motion stability and accuracy of the driver in the form of repeated practice; the monitoring takeover includes reminding the driver to observe the traffic environment in the process from the application of the driver to transfer the takeover control to the positive transfer of the takeover control; the auxiliary information includes: marking the road conditions around the vehicle through the screen; the signal optimization includes: prompting the driver by at least one of sound, light and vibration;

[0061] Before correction, the target regulation matrix M of the driver in any takeover scenario is generated according to the test in step 1:

[0062]

[0063] And after the correction, the updated target regulation matrix M' of the driver in the above takeover scenario is generated according to the test in step 1:

[0064]

[0065] And the weight vector R of the to-be-regulated parameters of any takeover scenario is determined according to the distance between the optimal and the worst solution, and R=(r1, r2,..., r n ) T , and the formula (5) is used to calculate the coefficient of variation of a single parameter C kn :

[0066]

[0067] And when C kn is a positive index and increases after correction, C kn takes a positive value; and decreases after correction, C kn takes a negative value; and when C kn is a negative index and decreases after correction, C kn takes a positive value; and increases after correction, C kn takes a negative value; the coefficient of variation matrix C is obtained to represent the improvement effect of any takeover scenario:

[0068] The improvement effect Y of any takeover scenario is represented by formula (6):

[0069] Y=C·R (6)

[0070] When Y≥N, the correction method is used for correction of all scenarios, where N is a constant and is determined by the designer, N in the embodiment is 1-5, specifically, N in the embodiment is 4.

[0071] Step 4, scene complexity determination: for the test in step 1, the operation mode of any one scene is divided into only braking, only steering, braking first and then steering, steering first and then braking, and braking and steering simultaneously, a training sample set is established, and the index complexity is normalized to obtain the difficulty representation coefficient of each operation mode in the interval of 0 to 1, and the weight coefficient d of the scene is determined according to the difficulty representation coefficient;

[0072] Step 5, driver takeover performance evaluation: on the basis of the test in step 1, the formula (4) is used to evaluate the takeover performance E1 corresponding to the specific takeover behavior of the driver:

[0073]

[0074] And after at least completing the intervention of step 3, the above formula is used to calculate the takeover performance E2 after the intervention, and after E2 is greater than E1, it is proved that the control means is suitable for the driver, and the control means can be continuously used to continuously improve the takeover success rate of the driver.

[0075] Embodiment 2

[0076] Step 3, when determining the takeover performance, the mean value of all takeover performances of the driver under the same scene is processed to obtain the mean value of the takeover performance evaluation result E1', which is taken as the takeover performance corresponding to the specific takeover behavior of the driver in step 3.

[0077] The above is only an embodiment of the present application, and the specific technical solutions and / or common knowledge of the scheme are not described in detail. It should be noted that for those skilled in the art, without departing from the technical solutions of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application, which will not affect the effect and practicality of the present application. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions, characterized by: Includes the following steps: Step 1, Establish successful takeover samples: Determine the optimal takeover budget time and cumulative takeover success rate curve for the driver in any takeover scenario through pre-testing, and construct a set of mapping parameters for different takeover behaviors and the entire takeover process in that takeover scenario; Step 2, Driver Takeover Competency Prediction: Dynamically collect driver data including head and eye movements, limb positions, psychological load, vehicle motion parameters, and driving environment state parameters. Optimize these parameters using the mayfly algorithm and construct a real-time takeover state vector T. The Mahalanobis distance D between the real-time takeover state vector and the perfect takeover state vector is calculated using formula (1). M (S): （1） The Mahalanobis distance D between the real-time takeover state vector and the failed takeover state vector is calculated using formula (2). M (F): （2） a represents the dimension of the performance mapping parameter for this scenario, R is the representation parameter, μ is the mean of the real-time state vector, and P is the corresponding mapping parameter; The basic trust probability assignment m(s) for a perfect takeover and the basic trust probability assignment m(f) for a failed takeover are calculated using formula (3): （3） And set the decision threshold for assigning competence to ε, predicting that the driver can successfully take over when m(s)-m(f)>ε; Step 3, takeover performance intervention and regulation: When m(s)-m(f)≤ε in Step 2, correction is performed from at least one dimension, such as reinforcement training, monitoring takeover, auxiliary information and signal optimization. The reinforcement training includes strengthening at least one of the driver's visual search, psychological load, reaction time, action stability and accuracy through repeated practice. The monitoring takeover includes reminding the driver to observe the traffic environment during the process from the driver's application to relinquish control to the formal relinquishment of control. The auxiliary information includes: displaying the road conditions around the vehicle on the screen; the signal optimization includes: providing prompts to the driver using at least one of sound, light, and vibration.

2. The method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions as described in claim 1, characterized in that: Step 2: When m(s)-m(f)≤ε, send a takeover signal to the driver and slide to the next time window to repeat the operation of Step 2 to continue prediction. When m(s)-m(f)>ε, predict that the driver can take over successfully.

3. The method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions as described in claim 2, characterized in that: Step 1: Determine the limit take-off time during the test; Step 2: When the prediction continues, if the driver's take-off fails when sliding to the limit take-off time while still satisfying m(s)-m(f)≤ε, then the driver's take-off is predicted to fail.

4. The method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions as described in claim 3, characterized in that: Each takeover result is included in the takeover sample of step 1.

5. The method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions according to claim 1, characterized in that: It also includes step 4, scene complexity determination. For the experiment in step 1, the operation mode of any scene is divided into braking only, steering only, braking first and then steering, steering first and then braking, and braking and steering simultaneously. A training sample set is established, and the index complexity is normalized to obtain the difficulty representation coefficient of each operation mode mapped to the interval between 0 and 1. The weight coefficient d of the scene is determined according to the difficulty representation coefficient. Step 5, Driver takeover performance evaluation: Based on the experiment in Step 1, the takeover performance E1 corresponding to a specific takeover behavior of the driver is evaluated using formula (4): （4） Where Q represents the characteristic value of each parameter corresponding to the driver in this scenario; After at least one intervention in step 3 is completed, the post-intervention takeover performance E2 is calculated using the formula described above.

6. The method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions as described in claim 1, characterized in that: Step 3: Based on the experiments in Step 1, generate the target control matrix M for the driver under any takeover scenario: After completing the calibration, the updated target control matrix M for the driver under any takeover scenario is generated based on the experiment in step 1. , : And determine the weight vector R of the parameters to be controlled in any takeover scenario based on the superior-inferior solution distance method, where R = (r1, r2, ..., r...). n ) T The coefficient of variation of a single parameter, C, is calculated using formula (5). kn : （5） And when C kn If it is a positive indicator and increases after correction, then C kn Take a positive value; and decrease it after correction, then C kn Take a negative value; and when C kn If it is a negative indicator and decreases after correction, then C kn Take a positive value; and increase it after correction, then C kn Take the negative value; this yields the coefficient of variation matrix C, which represents the improvement effect for any takeover scenario: Formula (6) represents the performance improvement effect Y of any one of the takeover scenarios: （6） When Y≥N, this correction method is applied to the correction of all scenarios; Where k is the number of takeover scenarios, n is the number of parameters to be adjusted, and J is the value of a takeover performance characterization parameter under a specified takeover scenario.

7. The method for predicting the success rate of takeover and evaluating the intervention effect under human-machine co-driving conditions as described in claim 5, characterized in that: Step 5: When determining takeover performance, average all the driver's takeover performance in the same scenario to obtain the averaged takeover performance evaluation result E1'. E1' is used as the takeover performance corresponding to a specific takeover behavior of the driver in Step 3.

Images