Abnormal motion detection closed loop iterative optimization method, system, medium, and apparatus

Abstract

The application discloses an abnormal action detection closed-loop iterative optimization method, system, medium and equipment, and belongs to the technical field of data optimization. The method comprises the following steps: in a vehicle terminal, a vehicle model performs rough screening on collected driver face video data, and a large amount of low-precision video data is obtained; the large amount of low-precision video data is classified in a cloud server, and first normal data, first abnormal data and suspicious data are obtained; the suspicious data is accurately classified into second normal data and second abnormal data; the first normal data, the first abnormal data, the second normal data and the second abnormal data are used for model training, and an optimized vehicle model and an optimized judgment model are obtained and updated. According to the application, the data is collected in a high-recall-rate low-precision manner, the collected data is processed by using a model on the cloud server, the data processing accuracy is improved, and the data classification is more and more accurate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data optimization, in particular to an abnormal action detection closed-loop iterative optimization method, system, medium and equipment. BACKGROUND

[0002] During the driving of a vehicle, the driver often has some dangerous driving actions or behaviors. For example, smoking, drinking, yawning, making a phone call and closing eyes during driving. Therefore, accurately detecting the driver's behavior is of great importance to safe driving. The conventional method in the prior art is to collect driver data by using a low-recall high-precision data collection method at the vehicle end, and to collect and classify the data of the driver's behavior at the vehicle end. The data that cannot be accurately analyzed in the low-recall high-precision data is reconfirmed by manual labeling. Since the data collected by the low-recall high-precision data collection method is of high precision, the manual reconfirmation process is redundant. In addition, the low-recall high-precision data collection method makes the amount of collected data very small, and many data containing various behavior information are filtered out under the low-recall high-precision collection method, so that the model cannot learn more content when further optimizing the model, and the improvement space of the model is small. SUMMARY

[0003] In view of the problem in the prior art that when collecting and processing the data of the driver's behavior, the vehicle terminal adopts a high-precision low-recall data collection method, which filters out data containing a large amount of information, resulting in a small amount of available data and the model cannot be further optimized by using the data, the present application provides an abnormal action detection closed-loop iterative optimization method, system, medium and equipment.

[0004] In one technical solution of the present application, an abnormal action detection closed-loop iterative optimization method is provided, comprising: in the vehicle terminal, the vehicle model performs rough screening on the collected driver face video data to obtain a large amount of low-precision video data; in the cloud server, the judgment model classifies the large amount of low-precision video data to obtain first normal data, first abnormal data and suspicious data; the suspicious data is accurately classified into second normal data and second abnormal data; the first normal data, the first abnormal data, the second normal data and the second abnormal data are used for model training to obtain an optimized vehicle model and an optimized judgment model, and the vehicle model and the judgment model are updated respectively.

[0005] Optionally, the vehicle-mounted model performs rough screening on the collected driver face video data to obtain massive low-precision video data, including: analyzing the face data by using the vehicle-mounted model to determine a first probability that the face data belongs to normal data and a second probability that the face data belongs to abnormal data; if the first probability or the second probability is greater than a corresponding first preset threshold, the face data is taken as the massive low-precision video data and uploaded to the cloud server.

[0006] Optionally, the massive low-precision video data is classified by a judgment model in the cloud server to obtain first normal data, first abnormal data and suspicious data, including: performing window processing on the massive low-precision video data by using a large model to obtain a plurality of window data; analyzing each window data to determine a third probability that each window data belongs to normal data and a fourth probability that each window data belongs to abnormal data; calculating the mean values of the plurality of third probabilities and the fourth probabilities, respectively; if the mean value of the third probabilities is greater than the mean value of the fourth probabilities, the massive low-precision video data is determined as the first normal data or the first abnormal data, otherwise, the massive low-precision video data is determined as the suspicious data.

[0007] Optionally, the massive low-precision video data is classified by a judgment model in the cloud server to obtain first normal data, first abnormal data and suspicious data, and further including: analyzing the plurality of window data by using a time sequence model to determine a fifth probability that the massive low-precision video data belongs to normal data or abnormal data in the time dimension; judging the massive low-precision video data according to the fifth probability, the mean value of the third probabilities and the mean value of the fourth probabilities to determine the massive low-precision video data as the first normal data, the first abnormal data or the suspicious data.

[0008] Optionally, the massive low-precision video data is classified by a judgment model in the cloud server to obtain first normal data, first abnormal data and suspicious data, and further including: detecting the posture of the driver in the massive low-precision video data by using an auxiliary detection model to obtain a judgment score corresponding to the massive low-precision video data; judging the massive low-precision video data according to the judgment score, the fifth probability, the mean value of the third probabilities and the mean value of the fourth probabilities to determine the massive low-precision video data as the first normal data, the first abnormal data or the suspicious data.

[0009] Optionally, the first normal data, the first abnormal data, the second normal data and the second abnormal data are used for model training to obtain an optimized vehicle-mounted model and an optimized judgment model, including: performing model training on the massive first normal data, the first abnormal data, the second normal data and the second abnormal data to optimize the optimized vehicle-mounted model and the optimized judgment model, respectively.

[0010] Optionally, the method further comprises: performing data evaluation by using the first normal data, the first abnormal data, the second normal data and the second abnormal data, so that comprehensive evaluation results of the driver driving can be obtained.

[0011] In one of the technical solutions of the present application, an abnormal action detection closed-loop iterative optimization system is provided, which comprises: a vehicle terminal, a vehicle model of the vehicle terminal performs rough screening on collected driver face video data to obtain massive low-precision video data; a cloud server, the cloud server classifies the massive low-precision video data by using a judgment model to obtain first normal data, first abnormal data and suspicious data, and performs accurate classification on the suspicious data to classify the suspicious data into second normal data and second abnormal data; and a post-processing module, the post-processing module performs model training by using the first normal data, the first abnormal data, the second normal data and the second abnormal data to obtain an optimized vehicle model and an optimized judgment model, and updates the vehicle model and the judgment model respectively.

[0012] In one of the technical solutions of the present application, a computer readable storage medium is provided, the storage medium stores computer instructions, and the computer instructions are operated to execute the method in the first solution.

[0013] In one of the technical solutions of the present application, a computer device is provided, which comprises a processor and a memory, and the memory stores computer instructions, wherein: the processor operates the computer instructions to execute the method in the first solution.

[0014] The present application has the following beneficial effects: the present application collects data by using a data collection method with high recall rate and low precision, obtains a large amount of sufficient data containing driver behavior information, processes the collected data by using a model on a cloud server, improves the precision of data processing, optimizes the model by using the data, and makes the optimized model more accurate in data processing, so that the classification of data is more and more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0016] Figure 1 is a flowchart of an embodiment of the abnormal action detection closed-loop iterative optimization method of the present application;

[0017] Figure 2 is a structural schematic diagram of an embodiment of the abnormal action detection closed-loop iterative optimization system of the present application;

[0018] Figure 3 is a structural schematic diagram of one example of the abnormal action detection closed loop iterative optimization system of the present application.

[0019] The specific embodiments of the present application have been shown through the above-described drawings, and will be described in more detail hereinafter. The drawings and the written description are not intended to restrict the scope of the present application in any way, but to explain the present application to those skilled in the art by referring to a specific embodiment. DETAILED DESCRIPTION

[0020] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without any creative work fall within the scope of protection of the present application.

[0021] The terms "first", "second", "third", "fourth" and the like (if any) in the description and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a product or device comprising a series of steps or units does not necessarily have to be limited to the units clearly listed, but can include other units not clearly listed or inherent to the product or device.

[0022] During driving, the driver often has some dangerous driving actions or behaviors. For example, smoking, drinking, yawning, making a phone call, and closing eyes during driving. Therefore, accurate detection of driver behavior is of great importance for safe driving. The conventional method in the prior art is to collect driver data at the vehicle end by using a low-recall high-precision data collection method, and to complete data collection and driver behavior classification processing at the vehicle end. The data that cannot be accurately analyzed in the low-recall high-precision data is reconfirmed by manual annotation. Since the data collected by the low-recall high-precision data collection method has high accuracy, the manual reconfirmation process is redundant. In addition, the low-recall high-precision data collection method makes the amount of collected data very small, and many data containing various behavior information are filtered out under the low-recall high-precision collection method, so that the model cannot learn more content when further optimizing the model, and the improvement space of the model is small.

[0023] To solve the above problems, the present application provides an abnormal action detection closed-loop iterative optimization method, system, medium and equipment. The method comprises: in the vehicle terminal, the vehicle model performs rough screening on the collected driver face video data to obtain a large amount of low-precision video data; in the cloud server, the first normal data, the first abnormal data and the suspicious data are obtained by classifying the large amount of low-precision video data by using the judgment model; the suspicious data is accurately classified into the second normal data and the second abnormal data; the first normal data, the first abnormal data, the second normal data and the second abnormal data are used for model training to obtain an optimized vehicle model and an optimized judgment model, and the vehicle model and the judgment model are updated respectively.

[0024] The abnormal action detection closed-loop iterative optimization method of the present application collects data by using a high-recall low-precision data collection method, obtains a large amount of sufficient data containing driver behavior information, processes the collected data by using the model on the cloud server, improves the accuracy of data processing, and optimizes the model by using the data, so that the optimized model can more accurately process data, and the classification of data is more and more accurate.

[0025] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0026] Figure 1 An embodiment of the abnormal action detection closed-loop iterative optimization method of the present application is shown.

[0027] In Figure 1 In the embodiment shown, the abnormal action detection closed-loop iterative optimization method of the application comprises a process S101, in the vehicle terminal, the vehicle model coarsely screens the collected driver face video data to obtain a large amount of low-precision video data.

[0028] In this embodiment, the abnormal actions of the driver targeted by the application mainly include smoking, drinking, yawning, closing eyes, and making a phone call, etc. When collecting data, the driver can be photographed by the vehicle camera to obtain the behavior data of the driver. Then the vehicle model is used to analyze the data photographed by the camera to determine whether the driver photographed by the camera has the above-mentioned abnormal actions. When performing specific analysis, the picture photographed by the camera can be cropped to obtain the face data of the driver, and the data is analyzed to obtain the probability that the data belongs to abnormal data, and the probability value is used for coarse screening to finally obtain a large amount of low-precision video data.

[0029] Optionally, the vehicle model coarsely screens the collected driver face video data to obtain a large amount of low-precision video data, comprising: using the vehicle model to analyze the face data to determine a first probability that the face data belongs to normal data and a second probability that the face data belongs to abnormal data; if the first probability or the second probability is greater than a corresponding first preset threshold, the face data is taken as the low-precision video data and uploaded to the cloud server.

[0030] In this optional embodiment, the vehicle model is used to analyze the face data to determine a first probability that the face video data belongs to normal data and a second probability that the face video data belongs to abnormal data. That is, whether the driver has the behaviors of smoking, drinking, etc. is analyzed, and the corresponding first probability and second probability are determined. By comparing the relationship between the first probability, the second probability and the corresponding first preset threshold, if the first probability or the second probability is greater than the first preset threshold, the face video data is taken as the low-precision video data and uploaded to the server. In this application, a low-precision high-recall data collection method is adopted, and the first preset threshold generally has a low value to obtain more data. It should be noted that the value of the first preset threshold can be reasonably set according to actual requirements, and the application does not make specific limitations.

[0031] Specifically, for example, when analyzing the driver's face data in three frames of pictures respectively, the probability of the first frame belonging to normal data is 40%, the probability of the second frame belonging to abnormal data is 50%, the probability of the third frame data normal data is 20%, and the probability of abnormal data is 22%. Assuming that the value of the first preset threshold is 30%, after judgment, the face video data corresponding to the first frame and the face video data corresponding to the second frame can be uploaded to the cloud server as low-precision video data, and the face video data corresponding to the third frame is filtered out. Compared with the high-precision low-recall rate mode in the prior art, which often sets a high threshold, if the first preset threshold is set to 45%, only the face video data corresponding to the second frame can be uploaded to the cloud server as low-precision video data. The data collection method in the prior art results in a small amount of data and difficulty in recalling difficult data, which is not convenient for subsequent operation and model optimization.

[0032] In Figure 1 In the embodiment shown, the abnormal action detection closed-loop iterative optimization method of the application includes a process S102, in which the cloud server classifies a large amount of low-precision video data through a judgment model to obtain first normal data, first abnormal data, and suspicious data.

[0033] In this embodiment, the cloud server can load a more powerful judgment model to further analyze and classify a large amount of low-precision video data, and subdivide the large amount of low-precision video data into normal data, first abnormal data, and suspicious data. The suspicious data is a large amount of low-precision video data that cannot be clearly determined as normal data or abnormal data after being judged by the model, so this part of data is treated as suspicious data.

[0034] Optionally, the cloud server classifies a large amount of low-precision video data through a judgment model to obtain first normal data, first abnormal data, and suspicious data, including: using a large model to perform sliding window processing on a large amount of low-precision video data to obtain a plurality of window data; analyzing each window data to determine a third probability of belonging to normal data and a fourth probability of belonging to abnormal data; calculating the mean of the plurality of third probabilities and fourth probabilities, respectively; if the mean of the third probabilities or the mean of the fourth probabilities is greater than the corresponding second preset threshold, the large amount of low-precision video data is determined as first normal data or first abnormal data, otherwise the large amount of low-precision video data is determined as suspicious data.

[0035] In the optional embodiment, in the cloud server, the screened video data is further analyzed by a large model with strong processing capability. The screened video data is first subjected to sliding window processing to obtain a plurality of window data. The plurality of window data are analyzed respectively to determine a third probability that each window data belongs to normal data and a fourth probability that each window data belongs to abnormal data. The mean values of the plurality of third probabilities and fourth probabilities are calculated, and the relationship between the mean values of the third probabilities or the mean values of the fourth probabilities and the second preset threshold is determined. If the mean value of the third probabilities is greater than the corresponding second preset threshold, the low-precision video data is determined as first normal data; if the mean value of the fourth probabilities is greater than the corresponding second preset threshold, the low-precision video data is determined as first abnormal data; otherwise, the low-precision video data is determined as suspicious data.

[0036] Specifically, the screened video data can be a 5-second video, and the plurality of window data obtained after sliding window processing can be 2-second video data. After analyzing the 2-second window data, a second probability of the window data is obtained. Table 1 shows an example of the second probability of the window data as follows:

[0037] Window data 1 Window data 2 Window data 3 Mean of probabilities First normal data 55％ 60％ 65％ 60％ First abnormal data 30％ 15％ 25* 35％

[0038] As shown in the above table, in the calculation of the probability mean value, the mean value of the third probability belonging to the first normal data is 60%, and the mean value of the fourth probability belonging to the first abnormal data is 35%. If the second preset threshold is set to 50%, it can be determined that the screened video data belongs to the first normal data. It should be noted that the above settings are only examples for illustrating the principles of the present application, and the specific values can be reasonably selected according to actual requirements.

[0039] Optionally, the large amount of low-precision video data is classified by a judgment model in the cloud server to obtain first normal data, first abnormal data and suspicious data, and further comprising: analyzing the plurality of window data by using a time sequence model to determine a fifth probability that the large amount of low-precision video data belongs to normal data or abnormal data in the time dimension; and judging the large amount of low-precision video data according to the fifth probability, the mean value of the third probability and the mean value of the fourth probability to determine that the large amount of low-precision video data is first normal data, first abnormal data or suspicious data.

[0040] In the optional embodiment, in order to improve the accuracy of data analysis, the present application uses a time sequence model in the cloud server to analyze the plurality of window data to determine a fifth probability that the large amount of low-precision video data belongs to normal data or abnormal data in the time dimension. Then, the fifth probability, the mean value of the third probability and the mean value of the fourth probability analyzed by the large model are comprehensively analyzed to determine that the low-precision video data is first normal data, first abnormal data or suspicious data.

[0041] Specifically, in the comprehensive judgment of the fifth probability, the third probability mean value and the fourth probability mean value, different weights can be allocated for comprehensive consideration, or other ways can be taken. The final determination result is more accurate.

[0042] Optionally, the cloud server classifies the massive low-precision video data through the judgment model to obtain the first normal data, the first abnormal data and the suspicious data, and further comprises: detecting the posture of the driver in the massive low-precision video data by using an auxiliary detection model to obtain a judgment score corresponding to the massive low-precision video data; and judging the massive low-precision video data according to the judgment score, the fifth probability, the third probability mean value and the fourth probability mean value to determine that the screening video data is the first normal data, the first abnormal data or the suspicious data.

[0043] In this embodiment, in an actual scene, the driver will have some behaviors such as lowering his head to play a mobile phone. At this time, the eyes of the driver are blocked, and since the model does not detect the eyes in the process of analysis and judgment, an incorrect judgment result may occur. Therefore, for the above special situation, the application detects the posture of the driver in the screening video data by using an auxiliary detection model to obtain a judgment score corresponding to the screening video data. When the auxiliary detection model specifically judges the posture of the driver, it judges the rotation angle of the head of the driver, the distance between the upper eyelid and the lower eyelid of the eyes, etc. to comprehensively judge whether the behavior of the driver belongs to normal action or abnormal action, and respectively obtains the judgment score. Through the comprehensive judgment of the judgment score, the fifth probability, the third probability mean value and the fourth probability mean value on the massive low-precision video data, the low-precision video data is finally determined to be the first normal data, the first abnormal data or the suspicious data, and the detection accuracy is improved.

[0044] In Figure 1 In the embodiment shown, the abnormal action detection closed-loop iterative optimization method of the application comprises process S103, which classifies the suspicious data accurately and classifies the suspicious data into second normal data and second abnormal data.

[0045] In this embodiment, for the suspicious data in the massive low-precision video data that cannot be classified by the model, the suspicious data is classified into second normal data and second abnormal data by manual labeling.

[0046] In Figure 1 In the embodiment shown, the abnormal action detection closed-loop iterative optimization method of the application comprises process S104, which uses the first normal data, the first abnormal data, the second normal data and the second abnormal data to train the model to obtain an optimized vehicle-mounted model and an optimized judgment model, and updates the vehicle-mounted model and the judgment model respectively.

[0047] Optionally, the first normal data, the first abnormal data, the second normal data, and the second abnormal data are used for model training to obtain an optimized vehicle-mounted model and an optimized judgment model, including: using a large amount of the first normal data, the first abnormal data, the second normal data, and the second abnormal data for model training, and respectively optimizing the optimized vehicle-mounted model and the optimized judgment model.

[0048] In the optional embodiment, a large amount of and classified accurate first normal data, first abnormal data, second normal data, and second abnormal data are obtained through the above process. By training the model using the accurate data, the model trained using the data can better fit the model function, and the accuracy of the model in processing data is also higher.

[0049] Optionally, the abnormal action detection closed-loop iterative optimization method of the application further includes: using the first normal data, the first abnormal data, the second normal data, and the second abnormal data for data evaluation, so that comprehensive evaluation results of the driver driving can be obtained.

[0050] In the optional embodiment, the first normal data, the first abnormal data, the second normal data, and the second abnormal data are also used for data evaluation. Since the low-precision high-recall rate data collection method is used, the amount of data is greatly improved, and various situations can be perfectly covered, so that the evaluation process is more comprehensive.

[0051] The abnormal action detection closed-loop iterative optimization method of the application collects data by using the high-recall rate low-precision data collection method, obtains sufficient low-precision video data containing driver behavior information, processes the collected data by using the model on the cloud server, improves the data processing precision, optimizes the model using the data, so that the optimized model can more accurately process data, and the data classification is more and more accurate. Through high-recall rate low-precision data collection and high-recall rate high-precision data processing, more usable data is finally obtained. Through model training and optimization, the whole process forms a closed loop, and is continuously optimized, so that the collected data is more and more in line with the requirements, the data is more and more clean, the model is more and more optimized, and the data that cannot be detected and judged before can be completely judged by the model after optimization, so that the manual annotation process is saved, and the model data processing capability is improved.

[0052] Figure 2 An embodiment of the abnormal action detection closed-loop iterative optimization system of the application is shown.

[0053] In Figure 2In the illustrated embodiment, the abnormal action detection closed-loop iterative optimization system of the present application comprises: a vehicle terminal 201, which performs rough screening on the collected driver face video data through a vehicle model to obtain massive low-precision video data; a cloud server 202, which classifies the massive low-precision video data through a judgment model to obtain first normal data, first abnormal data, and suspicious data, and performs accurate classification on the suspicious data to classify the suspicious data into second normal data and second abnormal data; and a post-processing module 203, which performs model training on the first normal data, the first abnormal data, the second normal data, and the second abnormal data to obtain an optimized vehicle model and an optimized judgment model, and updates the vehicle model and the judgment model respectively.

[0054] Optionally, in the vehicle terminal 201, the face data is analyzed by using the vehicle model to determine a first probability that the face data belongs to normal data and a second probability that the face data belongs to abnormal data; if the first probability or the second probability is greater than a corresponding first preset threshold, the face data is taken as the massive low-precision video data and uploaded to the cloud server.

[0055] Optionally, in the cloud server 202, the massive low-precision video data is subjected to sliding window processing by using a large model to obtain a plurality of window data; each window data is analyzed to determine a third probability that the window data belongs to normal data and a fourth probability that the window data belongs to abnormal data; the average values of the plurality of third probabilities and the fourth probabilities are calculated respectively, and if the average value of the third probabilities is greater than the average value of the fourth probabilities, the massive low-precision video data is determined as the first normal data or the first abnormal data, otherwise, the massive low-precision video data is determined as the suspicious data.

[0056] Optionally, in the cloud server 202, the plurality of window data is analyzed by using a time series model to determine a fifth probability that the massive low-precision video data belongs to normal data or abnormal data in the time dimension; the massive low-precision video data is judged according to the fifth probability, the average value of the third probabilities, and the average value of the fourth probabilities to determine the massive low-precision video data as the first normal data, the first abnormal data, or the suspicious data.

[0057] Optionally, in the cloud server 202, the posture of the driver in the massive low-precision video data is detected by using an auxiliary detection model to obtain a judgment score corresponding to the massive low-precision video data; the massive low-precision video data is judged according to the judgment score, the fifth probability, the average value of the third probabilities, and the average value of the fourth probabilities to determine the screening video data as the first normal data, the first abnormal data, or the suspicious data.

[0058] Figure 3 An example of the abnormal action detection closed-loop iterative optimization system of the present application is shown.

[0059] AsFigure 3 As shown, first, each vehicle-mounted terminal collects data in a high-recall low-precision manner, and then transmits the data to a cloud server. On the cloud server, large models, sequence models, and auxiliary detection models are used to process the data. For data that cannot be judged by some models, artificial marking methods are used for judgment. The final data is evaluated and the model is trained, so that the model is more and more optimized, and the data that cannot be detected and judged before can also be completely judged by the model after optimization. Finally, the artificial marking process is saved.

[0060] The abnormal action detection closed-loop iterative optimization system of the present application collects data by using a high-recall low-precision data collection method, obtains a large amount of sufficient data containing driver behavior information, and processes the collected data using models on a cloud server to improve the accuracy of data processing. The data is used to optimize the model, so that the optimized model can more accurately process data, and the classification of data becomes more and more accurate.

[0061] In one specific embodiment of the present application, a computer-readable storage medium stores computer instructions, wherein the computer instructions are operated to perform the abnormal action detection closed-loop iterative optimization method described in any embodiment. Wherein the storage medium can be directly in hardware, in a software module executed by a processor, or in a combination of the two.

[0062] The software module can reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. The exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium.

[0063] The processor can be a central processing unit (CPU), a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof. The general purpose processor can be a microprocessor, but in the alternative, the processor can be any conventional processor, controller, microcontroller, or state machine. The processor can also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. The storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0064] In one embodiment of the present application, a computer device includes a processor and a memory, the memory storing computer instructions, wherein the processor operates the computer instructions to perform the abnormal action detection closed loop iterative optimization method described in any embodiment.

[0065] In the embodiments provided by the present application, it should be understood that the disclosed apparatus can be implemented in other manners. For example, the described apparatus embodiments are merely illustrative, and for example, division into units is merely a logical function division, and an actual implementation can be another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0066] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0067] The above merely illustrates the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structural transformation or direct or indirect application in other related technical fields by using the content of the present application specification and drawings are also included in the patent protection scope of the present application.

Claims

1. A closed-loop iterative optimization method for abnormal action detection, characterized in that, include: In the vehicle terminal, the vehicle model performs a coarse screening of the collected driver facial video data to obtain multiple low-precision video data. This coarse screening includes: analyzing the facial video data using the vehicle model to determine a first probability that the facial video data belongs to normal data and a second probability that it belongs to abnormal data; if the first probability or the second probability is greater than a corresponding first preset threshold, then the facial video data is considered as the low-precision video data and uploaded to a cloud server. In the cloud server, a judgment model is used to classify multiple low-precision video data to obtain first normal data, first abnormal data, and questionable data. The questionable data is precisely classified into second normal data and second abnormal data. The first normal data, the first abnormal data, the second normal data, and the second abnormal data are used to train the model to obtain an optimized vehicle model and an optimized judgment model, and the vehicle model and the judgment model are updated respectively.

2. The abnormal action detection closed-loop iterative optimization method according to claim 1, characterized in that, The process of classifying multiple low-precision video data sets using a judgment model on a cloud server to obtain first normal data, first abnormal data, and questionable data includes: The low-precision video data is processed using a large model to obtain multiple window data. Analyze the data in each window to determine the third probability that it belongs to normal data and the fourth probability that it belongs to abnormal data. Calculate the average of multiple third probabilities and fourth probabilities respectively. If the average of the third probability or the average of the fourth probability is greater than the corresponding second preset threshold, the low-precision video data is determined as the first normal data or the first abnormal data; otherwise, the low-precision video data is determined as the questionable data.

3. The abnormal action detection closed-loop iterative optimization method according to claim 2, characterized in that, The step of classifying multiple low-precision video data in the cloud server using a judgment model to obtain first normal data, first abnormal data, and questionable data also includes: By analyzing multiple window data using a time series model, the fifth probability of whether the low-precision video data belongs to normal data or abnormal data is determined in the time dimension; The low-precision video data is judged based on the fifth probability, the average of the third probability, and the average of the fourth probability to determine whether the low-precision video data is the first normal data, the first abnormal data, or the questionable data.

4. The abnormal action detection closed-loop iterative optimization method according to claim 3, characterized in that, The step of classifying the low-precision video data in the cloud server using a judgment model to obtain first normal data, first abnormal data, and questionable data also includes: The driver's posture in the low-precision video data is detected using an auxiliary detection model, and the judgment score corresponding to the low-precision video data is obtained. The low-precision video data is judged based on the judgment score, the fifth probability, the average of the third probability, and the average of the fourth probability to determine whether the selected video data is the first normal data, the first abnormal data, or the questionable data.

5. The abnormal action detection closed-loop iterative optimization method according to claim 1, characterized in that, The step of training the model using the first normal data, the first abnormal data, the second normal data, and the second abnormal data to obtain an optimized vehicle-mounted model and an optimized judgment model includes: The optimized vehicle model and the optimized judgment model are optimized by training the model using multiple sets of first normal data, first abnormal data, second normal data, and second abnormal data.

6. The abnormal action detection closed-loop iterative optimization method according to claim 1, characterized in that, Also includes: By using the first normal data, the first abnormal data, the second normal data, and the second abnormal data for data evaluation, a comprehensive evaluation result of the driver's driving can be obtained.

7. An abnormal action detection closed-loop iterative optimization system, characterized in that, include: The vehicle-mounted terminal uses its vehicle-mounted model to coarsely screen collected driver facial video data, obtaining multiple low-precision video data. This coarse screening includes: analyzing the facial video data using the vehicle-mounted model to determine a first probability that the facial video data belongs to normal data and a second probability that it belongs to abnormal data; if either the first probability or the second probability is greater than a corresponding first preset threshold, the facial video data is considered as the low-precision video data and uploaded to a cloud server. The cloud server classifies multiple low-precision video data through a judgment model to obtain first normal data, first abnormal data and questionable data, and accurately classifies the questionable data into second normal data and second abnormal data. The post-processing module uses the first normal data, the first abnormal data, the second normal data, and the second abnormal data to train the model, obtain an optimized vehicle model and an optimized judgment model, and updates the vehicle model and the judgment model respectively.

8. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions that are operated to execute the abnormal action detection closed-loop iterative optimization method according to any one of claims 1-6.

9. A computer device comprising a processor and a memory, the memory storing computer instructions, wherein: The processor operates computer instructions to execute the abnormal action detection closed-loop iterative optimization method as described in any one of claims 1-6.

Images