Artificial intelligence-based traffic accident prediction method, device, equipment and medium

Abstract

This application provides an artificial intelligence-based traffic accident prediction method, apparatus, device, and medium. The method acquires images of accident vehicles, calls a vehicle type recognition sub-model in a trained traffic accident prediction model to identify the accident vehicle images, and obtains vehicle type labels. The traffic accident prediction model includes a vehicle type recognition sub-model and multiple prediction sub-models. The prediction sub-model corresponding to the vehicle type label is called to predict the accident vehicle images and obtain prediction results, thus realizing the prediction of accident vehicle images. Since the prediction sub-models that perform the prediction correspond one-to-one with different vehicle type labels, the prediction sub-models are more targeted for accident vehicle images with determined vehicle type labels, thereby improving the accuracy of accident vehicle image prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a traffic accident prediction method and device based on artificial intelligence, equipment and medium. BACKGROUND

[0002] With the rapid development of economy and technology, there are more and more different types of vehicles, such as cars, buses, trucks, electric vehicles, etc., but the number of traffic accidents has also increased accordingly.

[0003] At present, in the existing traffic accidents, it is usually identified by artificial on-site, which is time-consuming and laborious, resulting in traffic congestion, and there is a certain subjectivity, therefore, how to quickly and accurately predict traffic accidents is a problem to be solved at present. SUMMARY

[0004] The embodiments of the present application provide a traffic accident prediction method, device, equipment and medium based on artificial intelligence to solve the technical problem of low traffic accident prediction efficiency of artificial on-site identification.

[0005] In one aspect, the present application provides a traffic accident prediction method based on artificial intelligence, comprising:

[0006] obtaining an accident vehicle image;

[0007] calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image, and obtaining a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained according to a class activation mapping network model;

[0008] calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image, and obtaining a prediction result.

[0009] In one aspect, the present application provides a traffic accident prediction device based on artificial intelligence, comprising:

[0010] an acquisition module for acquiring an accident vehicle image;

[0011] an identification module for calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image, and obtaining a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained according to the framework of a class activation mapping network model;

[0012] The prediction module is configured to call the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image, and obtain a prediction result.

[0013] In one aspect, the present application provides a computer device, comprising a memory and a processor, the memory storing a computer program, and the computer program is executed by the processor to make the processor execute the following steps of the artificial intelligence-based traffic accident prediction method:

[0014] An accident vehicle image is obtained.

[0015] A vehicle type recognition sub-model in a trained traffic accident prediction model is called to recognize the accident vehicle image, and a vehicle type label is obtained, wherein the traffic accident prediction model comprises a vehicle type recognition sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained according to a class activation mapping network model.

[0016] The prediction sub-model corresponding to the vehicle type label is called to predict the accident vehicle image, and a prediction result is obtained.

[0017] In one aspect, the present application provides a computer readable medium storing a computer program, and the computer program is executed by a processor to make the processor execute the following steps of the artificial intelligence-based traffic accident prediction method:

[0018] An accident vehicle image is obtained.

[0019] A vehicle type recognition sub-model in a trained traffic accident prediction model is called to recognize the accident vehicle image, and a vehicle type label is obtained, wherein the traffic accident prediction model comprises a vehicle type recognition sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained according to a class activation mapping network model.

[0020] The prediction sub-model corresponding to the vehicle type label is called to predict the accident vehicle image, and a prediction result is obtained.

[0021] The embodiment of the application provides a traffic accident prediction method based on artificial intelligence, and the method. The application obtains an accident vehicle image, calls a vehicle type identification submodel in a trained traffic accident prediction model to identify the accident vehicle image, and obtains a vehicle type label, wherein the traffic accident prediction model comprises the vehicle type identification submodel and a plurality of prediction submodels, one vehicle type label corresponds to one prediction submodel, the prediction submodel is obtained according to a class activation mapping network model, a prediction submodel corresponding to the vehicle type label is called to predict the accident vehicle image, and a prediction result is obtained. The embodiment realizes prediction of the accident vehicle image. Since the prediction submodels for prediction correspond to different vehicle type labels one by one, the prediction submodel is more targeted for the accident vehicle image with the determined vehicle type label, the accuracy of prediction of the accident vehicle image is improved, and since each prediction submodel has interpretability, the features of the attention area for prediction are more accurate, and the accuracy of the prediction result is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the 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 only show some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0023] Among them:

[0024] Figure 1 It is a flow chart of the traffic accident prediction method based on artificial intelligence in one embodiment;

[0025] Figure 2 It is a schematic diagram of the traffic accident prediction model prediction process in one embodiment;

[0026] Figure 3 It is a schematic diagram of the first prediction submodel training process in one embodiment;

[0027] Figure 4 It is a structural block diagram of the traffic accident prediction device based on artificial intelligence in one embodiment;

[0028] Figure 5 It is a structural block diagram of the computer device in one embodiment. DETAILED DESCRIPTION

[0029] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0030] As shown in the figure, in one embodiment, an artificial intelligence-based traffic accident prediction method is provided, which can be applied to a terminal and a server. The embodiment is exemplified by application to a server. The artificial intelligence-based traffic accident prediction method specifically includes the following steps: Figure 1

[0031] Step 102, acquiring an accident vehicle image.

[0032] The accident vehicle image refers to a vehicle image of a traffic accident, which is used for subsequent identification of the type of traffic accident. Specifically, the accident vehicle image can be extracted from a video recorded by a driving recorder, can be collected on site by a camera device, or can be pre-stored in a server and obtained by calling from the server.

[0033] Step 104, calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image, and obtaining a vehicle type label, wherein the traffic accident prediction model includes a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is trained according to a class activation mapping network model.

[0034] ​The trained traffic accident prediction model includes a vehicle type identification sub-model and a plurality of prediction sub-models. The vehicle type identification sub-model is used to determine the vehicle type label of the accident vehicle image, such as a motor vehicle pair (motor vehicle-motor vehicle), a non-motor vehicle pair (non-motor vehicle-motor vehicle), and a non-motor vehicle pair (non-motor vehicle-non-motor vehicle). Specifically, the vehicle type identification sub-model can be a machine learning-based classifier, such as a support vector machine (SVM), a decision tree, a random forest, or the like. The prediction sub-model is used to predict the severity level of the traffic accident of the accident vehicle image with the determined vehicle type label. The plurality of prediction sub-models are provided, and one prediction sub-model corresponds to one vehicle type label, i.e., the number of prediction sub-models is consistent with the number of vehicle type labels. Each prediction sub-model is trained according to a class activation mapping (CAM) network model. The prediction sub-model in the embodiment can be improved based on the CAM model. The CAM model is an interpretable model. In order to obtain the sample area that contributes to the model prediction, the CAM model is used for interpretation, which can help people understand the basis for the decision of the machine learning model for each input sample, obtain a visual explanation result, and highlight the decision area for model prediction in the input sample. These attention areas provide insights into which information of the input sample is actually used for prediction by the model. For the problem that there is a gap between the image saliency area and the decision area, the CAM model provides better interpretability. It can be understood that in the embodiment, the prediction sub-model corresponding to each vehicle type is set, and the feature differences between the traffic accidents of different vehicle types are fully considered, so that the prediction sub-models are more targeted, which is beneficial to improve the prediction accuracy. Each prediction sub-model is trained based on the CAM model, so that the attention area of the prediction result, such as the decision area and the saliency area, is explained, the features of the attention area for prediction are more accurate, and the accuracy of the subsequent prediction result is improved.

[0035] In step 106, the prediction sub-model corresponding to the vehicle type label is called to predict the accident vehicle image, and a prediction result is obtained.

[0036] The prediction result is a classification result for reflecting the severity level of the traffic accident image of the accident vehicle. Specifically, for different types of accident vehicle images, corresponding prediction sub-models are used for predicting the severity level of the traffic accident. Since the prediction sub-models correspond one-to-one with different vehicle type labels, the prediction sub-models are more targeted for accident vehicle images with determined vehicle type labels, improving the accuracy of the prediction of the accident vehicle image. Moreover, since each prediction sub-model has interpretability, the features of the attention area for prediction are more accurate, further improving the accuracy of the prediction result. Compared with the traditional manual identification method, the prediction efficiency of the traffic accident is greatly improved.

[0037] The above-mentioned traffic accident prediction method based on artificial intelligence, by acquiring an accident vehicle image, calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image, obtaining a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, the prediction sub-model is trained according to a class activation mapping network model, the prediction sub-model corresponding to the vehicle type label is called to predict the accident vehicle image, and a prediction result is obtained. The embodiment realizes the prediction of the accident vehicle image. Since the prediction sub-models correspond one-to-one with different vehicle type labels, the prediction sub-models are more targeted for accident vehicle images with determined vehicle type labels, improving the accuracy of the prediction of the accident vehicle image. Moreover, since each prediction sub-model has interpretability, the features of the attention area for prediction are more accurate, further improving the accuracy of the prediction result.

[0038] In one embodiment, the vehicle type label includes a first label, a second label or a third label, and the prediction sub-model includes a first prediction sub-model corresponding to the first label, a second prediction sub-model corresponding to the second label or a third prediction sub-model corresponding to the third label. The step of calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image and obtaining the prediction result includes: if the vehicle type label is the first label, inputting the accident vehicle image into the first prediction sub-model for prediction to obtain the prediction result; if the vehicle type label is the second label, inputting the accident vehicle image into the second prediction sub-model for prediction to obtain the prediction result; and if the vehicle type label is the third label, inputting the accident vehicle image into the third prediction sub-model for traffic accident prediction to obtain the prediction result.

[0039] Specifically, the vehicle type label includes three types, i.e., a first label, a second label or a third label, and the prediction sub-models include three types, i.e., a first prediction sub-model corresponding to the first label, a second prediction sub-model corresponding to the second label or a third prediction sub-model corresponding to the third label. Specifically, after the vehicle type label is determined, the prediction sub-model corresponding to the vehicle type label is selected to predict the accident vehicle image, and the prediction result is determined according to the output of the selected prediction sub-model, as shown in FIG. 8, which is a schematic diagram of a traffic accident prediction model prediction process. Figure 2 The first prediction sub-model, the second prediction sub-model and the third prediction sub-model are all trained by using the CAM model, i.e., by using the CAM model or a corresponding variant CAM model such as Grad-CAM, Grad-CAM++, CALM (class activation latent mapping), F-CALM (Focus-class activation latent mapping) and the like. The CALM model can explicitly combine a latent variable that encodes the position of a recognition clue in the CAM model, so as to incorporate an attribution map into a training computational graph. In this way, the CALM model is constructed. For example, the position of each pixel point of the accident vehicle image can be taken as the latent variable, and the CALM model is trained by using an Expectation-Maximization algorithm (EM). Therefore, compared with Grad-CAM and Grad-CAM++, the CALM model can more accurately identify the discriminative attributes of an image classifier. The F-CALM model is a variant of the CALM model, which adds a function of focusing on an attention region. As a preferred example, the F-CALM model is selected for training, so as to make full use of the focusing function and the more accurate feature recognition capability of the F-CALM model to improve the accuracy of the features of the first prediction sub-model, the second prediction sub-model and the third prediction sub-model.

[0040] It is worth noting that the first, second and third prediction sub-models are trained and tested on different samples, and have different model parameters. The samples are determined according to the corresponding vehicle type label. For example, for the first prediction sub-model, the image of the accident vehicle with the first label can be selected as the sample for training and testing, and the parameters of each model are determined according to the training and testing results. Understandably, in the present embodiment, by learning the severity level of the traffic accident of different types of vehicles, the prediction sub-model corresponding to each vehicle type label is obtained, which makes the prediction of each prediction sub-model more fine-grained compared to a single model, and improves the adaptability and accuracy of each prediction sub-model for the prediction of the image of the accident vehicle corresponding to the vehicle type label.

[0041] In one embodiment, the first, second and third prediction sub-models are all focus-class activation latent mapping network models.

[0042] The focus-class activation latent mapping network model (F-CALM) is a model that adds a focus function for the attention area based on the CALM model. The focus function can add an area loss function to the CALM model to improve the focusing ability of the F-CALM model, achieve focusing on the attention area, and avoid the problem of deviation in the corresponding feature maps caused by the focusing area of the attention area being wrong when the line-of-sight focus point is wrong. The first, second and third prediction sub-models in the present embodiment are all F-CALM models. By using the focusing function in the F-CALM model, the line-of-sight focus area can still be displayed after the line-of-sight focus point is wrong, ensuring that the attention area does not completely lose focus, so that the features of the attention area are more accurate, thereby greatly improving the accuracy of the prediction of the first, second and third prediction sub-models.

[0043] In one embodiment, the training process of the first prediction sub-model includes: obtaining a training sample image set, the training sample image set including positive sample images with a vehicle type label of a first label and negative sample images with a vehicle type label of a second label or a third label; determining a preset class activation mapping network model, the loss function of the preset class activation mapping network model including an area loss function for learning the focusing ability of the preset class activation mapping network model; inputting the training sample image set into the class activation mapping network model for training, and generating the first prediction sub-model under the condition that the preset class activation mapping network model converges.

[0044] The preset class activation mapping network model can be one of a CAM, Grad-CAM, Grad-CAM++, CALM or F-CALM model, and the loss function of the preset class activation mapping network model includes an area loss function, which is used to learn the focusing ability of the preset class activation mapping network model. For the first prediction sub-model, when the model is trained, the accident vehicle image corresponding to the first label of the first prediction sub-model can be used as a positive sample image, and the accident vehicle images of the second label and the third label that do not match the vehicle type label of the first prediction sub-model can be used as negative sample images, so as to further improve the prediction accuracy of the first prediction sub-model for the accident vehicle image of the first label. Specifically, the training sample image set is input into the class activation mapping network model for training, and the parameters of the preset class activation mapping network model are optimized according to the loss function of the preset class activation mapping network model, the predicted value and the label value of the training sample. In the optimization process, when the loss value calculated according to the loss function of the preset class activation mapping network model is less than or equal to the preset loss standard value, it is determined that the trained preset class activation mapping network model converges, and thus the first prediction sub-model is generated.

[0045] It should be noted that the training processes of the second prediction sub-model and the third prediction sub-model are consistent with that of the first prediction sub-model, and the difference lies in that the training sample image set is different, that is, for the second prediction sub-model, the accident vehicle image corresponding to the second label of the second prediction sub-model can be used as a positive sample image, and the accident vehicle images of the first label and the third label that do not match the vehicle type label of the second prediction sub-model can be used as negative sample images; for the third prediction sub-model, the accident vehicle image corresponding to the third label of the third prediction sub-model can be used as a positive sample image, and the accident vehicle images of the second label and the third label that do not match the vehicle type label of the third prediction sub-model can be used as negative sample images, so as to improve the prediction accuracy of the respective models.

[0046] In one embodiment, the area loss function is:

[0047]

[0048] wherein z i is the position of the i-th pixel point, the number of which is N, x is the training image, is the true class label of the i-th pixel point.

[0049] Specifically, in the CALM model, the position z i of the i-th pixel point on the training image x is taken as a hidden variable, and the EM algorithm is used to learn the position z ithe conditional probability distribution of z given x. Assuming the size of the training image x is HxW, i∈{1, 2, …, HxW}, i.e. N = HxW, and the number of classes is C, the channel number of the feature map extracted by the classification network is changed to C and 1 through two convolutional layers respectively, and the softmax is performed on the feature map of CxHxW along the channel to obtain L1 normalization is performed on the feature map of 1xHxW to obtain h z = p(z|x), After broadcasting, it is multiplied element by element with to obtain the joint probability distribution p(y, z i |x). The feature map of the classification network is is the true class label of the training image x. Let p θ′ (z i |x, y) be the probability distribution of z given that the classification result of the input image x is y, then the loss function of the CALM model is obtained according to the EM algorithm as

[0050] L EM = -log p θ (y|x) ≤ -∑ i p θ′ (z i |x, y) log p θ (y, z i |x);

[0051] where θ is the parameter of the classification network, and θ' is the parameter of the distribution of z.

[0052]

[0053] Therefore, the total loss function of the predictor model is

[0054] L = L EM + λL area ;

[0055] where λ is a hyperparameter, and L is the total loss function of the predictor model.

[0056] In one embodiment, the first label is a pair of motor vehicles, the second label is a pair of motor vehicles and non-motor vehicles, and the third label is a pair of non-motor vehicles, wherein the motor vehicles include cars or trucks, the non-motor vehicles include electric vehicles or bicycles, the preset class activation mapping network model includes a first class activation mapping network sub-model, a second class activation mapping network sub-model, and a third class activation mapping network sub-model, and before the step of determining the preset class activation mapping network model, the method further includes: identifying the positive sample image, if the positive sample image is a car and a truck, inputting the training sample image set into the first class activation mapping network sub-model for training, and generating the first prediction sub-model under the condition that the first class activation mapping network sub-model converges; if the positive sample image is a car and a car, inputting the training sample image set into the second class activation mapping network sub-model for training, and generating the first prediction sub-model under the condition that the second class activation mapping network sub-model converges; and if the positive sample image is a truck and a truck, inputting the training sample image set into the third class activation mapping network sub-model for training, and generating the first prediction sub-model under the condition that the third class activation mapping network sub-model converges.

[0057] Specifically, after the vehicle type label is determined, the vehicle type contained in the vehicle type label is continuously identified to determine the vehicle type corresponding to the accident vehicle image, for example, whether it is a car or a truck, or whether it is a bicycle or an electric vehicle. For the first label, different vehicle types are trained by using the corresponding first class activation mapping network sub-model, second class activation mapping network sub-model, or third class activation mapping network sub-model, so as to further improve the adaptability and accuracy of each prediction sub-model in predicting the accident vehicle image of a specific vehicle type. As shown in FIG. 4, it is a schematic diagram of the training process of the first prediction sub-model. Figure 3

[0058] It should be noted that for the second label, the positive sample image includes four situations of a truck and a bicycle, a truck and an electric vehicle, a car and a bicycle, and a car and an electric vehicle. Then, the preset class activation mapping network model includes a fourth class activation mapping network sub-model, a fifth class activation mapping network sub-model, a sixth class activation mapping network sub-model, and a seventh class activation mapping network sub-model. According to the vehicle type in the positive sample image, the corresponding preset class activation mapping network model is selected to train the second preset sub-model. For the third label, the positive sample image includes three situations of a bicycle and a bicycle, an electric vehicle and an electric vehicle, and an electric vehicle and a bicycle. Then, the preset class activation mapping network model includes an eighth class activation mapping network sub-model, a ninth class activation mapping network sub-model, and a tenth class activation mapping network sub-model. According to the vehicle type in the positive sample image, the corresponding preset class activation mapping network model is selected to train the third preset sub-model.

[0059] ​In one embodiment, if the vehicle type label is a first label, the accident vehicle image is input into a first prediction sub-model for prediction to obtain a prediction result, including: extracting a saliency map and a corresponding feature map of the accident vehicle image by using the first prediction sub-model; determining the prediction result based on the saliency map and the feature map.

[0060] The saliency map refers to an attention area of the accident vehicle image, for example, an area of a collision of a traffic accident. Specifically, the saliency map of the accident vehicle image is extracted by F-CALM in the first prediction sub-model, the corresponding feature map is obtained by feature extraction on the saliency map, and the prediction result is obtained by analyzing the feature map. The prediction result can be reflected by an accident score or an accident level. Understandably, in this embodiment, the saliency map has explainability and focus, thereby improving the accuracy of the prediction result.

[0061] In one embodiment, after the step of calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image to obtain a prediction result, the method further includes: determining an accident level corresponding to the accident vehicle image according to the prediction result; and determining a claim strategy of the accident vehicle according to a preset mapping relationship table of the accident level and the claim strategy and the accident level.

[0062] Specifically, the preset mapping relationship table of the accident level and the claim strategy is stored in the server in advance, and the table records each accident level and a corresponding claim strategy. After the accident level is determined, the corresponding claim strategy is queried in the mapping relationship table, thereby realizing automatic claim. Since the prediction result in this embodiment is highly accurate, the claim strategy determined based on the prediction result is also more reasonable, thereby greatly improving the claim efficiency of vehicle insurance.

[0063] As shown in FIG. 1, Figure 4 In one embodiment, an artificial intelligence-based traffic accident prediction device is provided, which includes:

[0064] The acquisition module 402 is configured to acquire an accident vehicle image.

[0065] The identification module 404 is configured to call a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image to obtain a vehicle type label. The traffic accident prediction model includes the vehicle type identification sub-model and a plurality of prediction sub-models. One vehicle type label corresponds to one prediction sub-model. The prediction sub-model is trained according to a framework of a class activation mapping network model.

[0066] The prediction module 406 is configured to call the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image to obtain a prediction result.

[0067] In one embodiment, the prediction module comprises:

[0068] a first prediction unit, configured to input the accident vehicle image into the first prediction sub-model for prediction if the vehicle type label is the first label, to obtain the prediction result;

[0069] a second prediction unit, configured to input the accident vehicle image into the second prediction sub-model for prediction if the vehicle type label is the second label, to obtain the prediction result;

[0070] a third prediction unit, configured to input the accident vehicle image into the third prediction sub-model for traffic accident prediction if the vehicle type label is the third label, to obtain the prediction result.

[0071] In one embodiment, the artificial intelligence-based traffic accident prediction device further comprises:

[0072] a collection module, configured to obtain a training sample image set, the training sample image set comprising positive sample images with the vehicle type label being the first label and negative sample images with the vehicle type label being the second label or the third label;

[0073] a determination module, configured to determine a preset class activation mapping network model, a loss function of the preset class activation mapping network model comprising an area loss function, the area loss function being used to learn focusing ability of the preset class activation mapping network model;

[0074] a training module, configured to input the training sample image set into the class activation mapping network model for training, and generate a first prediction sub-model in a case where the preset class activation mapping network model converges.

[0075] In one embodiment, the artificial intelligence-based traffic accident prediction device further comprises:

[0076] a first training module, configured to identify the positive sample images, and input the training sample image set into the first class activation mapping network sub-model for training if the positive sample images are sedans and trucks, and generate a first prediction sub-model in a case where the first class activation mapping network sub-model converges;

[0077] a second training module, configured to input the training sample image set into the second class activation mapping network sub-model for training if the positive sample images are sedans and sedans, and generate a first prediction sub-model in a case where the second class activation mapping network sub-model converges;

[0078] The third training module is used to input the training sample image set into the third type activation mapping network sub-model for training if the positive sample images are trucks and trucks, and to generate the first prediction sub-model if the third type activation mapping network sub-model converges.

[0079] In one embodiment, the first prediction unit includes:

[0080] An extraction subunit is used to extract the saliency map and corresponding feature map of the accident vehicle image using the first prediction submodel;

[0081] A prediction subunit is used to determine the prediction result based on the saliency map and the feature map.

[0082] In one embodiment, the AI-based traffic accident prediction device further includes:

[0083] The first determining module is used to determine the accident level corresponding to the accident vehicle image based on the prediction result;

[0084] The second determining module is used to determine the claim strategy for the accident vehicle based on a preset mapping table between accident levels and claim strategies and the accident level.

[0085] Figure 5 An internal structural diagram of a computer device in one embodiment is shown. This computer device may specifically be a server, including but not limited to high-performance computers and high-performance computer clusters. Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an artificial intelligence-based traffic accident prediction method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the artificial intelligence-based traffic accident prediction method. Those skilled in the art will understand that... Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0086] In one embodiment, the AI-based traffic accident prediction method provided in this application can be implemented as a computer program, which can be implemented in the form of, for example... Figure 5The computer device shown is running. The memory of the computer device can store various program templates constituting the artificial intelligence-based traffic accident prediction device. For example, the acquisition module 402, the identification module 404, and the prediction module 406.

[0087] A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the following artificial intelligence-based traffic accident prediction method:

[0088] Acquiring an accident vehicle image;

[0089] Calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image to obtain a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained according to a class activation mapping network model.

[0090] Calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image to obtain a prediction result.

[0091] A computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the steps of the following artificial intelligence-based traffic accident prediction method:

[0092] Acquiring an accident vehicle image;

[0093] Calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image to obtain a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained according to a class activation mapping network model.

[0094] Calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image to obtain a prediction result.

[0095] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0097] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for predicting a traffic accident based on artificial intelligence, the method comprising: The method comprises: acquiring an accident vehicle image; calling a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image, to obtain a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is obtained by training a class activation mapping network model; calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image, to obtain a prediction result; wherein the vehicle type label comprises a first label, a second label, or a third label, the first label is a motor vehicle pair, the second label is a motor vehicle and non-motor vehicle pair, and the third label is a non-motor vehicle pair, a motor vehicle comprises a car or a truck, and a non-motor vehicle comprises an electric vehicle or a bicycle; the prediction sub-model comprises a first prediction sub-model corresponding to the first label, a second prediction sub-model corresponding to the second label, or a third prediction sub-model corresponding to the third label; if the vehicle type label is the first label, the accident vehicle image is input into the first prediction sub-model for prediction, to obtain the prediction result; if the vehicle type label is the second label, the accident vehicle image is input into the second prediction sub-model for prediction, to obtain the prediction result; and if the vehicle type label is the third label, the accident vehicle image is input into the third prediction sub-model for traffic accident prediction, to obtain the prediction result.

2. The artificial intelligence-based traffic accident prediction method of claim 1, wherein, The first prediction sub-model, the second prediction sub-model, and the third prediction sub-model are all focus-class activation latent mapping network models.

3. The traffic accident prediction method based on artificial intelligence according to claim 2, wherein the training process of the first prediction sub-model comprises: acquiring a training sample image set, wherein the training sample image set comprises positive sample images with the vehicle type label being the first label and negative sample images with the vehicle type label being the second label or the third label; determining a preset class activation mapping network model, wherein a loss function of the preset class activation mapping network model comprises an area loss function, and the area loss function is used to learn the focusing ability of the preset class activation mapping network model; inputting the training sample image set into the class activation mapping network model for training, and generating the first prediction sub-model when the preset class activation mapping network model converges.

4. The traffic accident prediction method based on artificial intelligence according to claim 3, wherein the preset class activation mapping network model comprises a first class activation mapping network sub-model, a second class activation mapping network sub-model, and a third class activation mapping network sub-model, and before the step of determining the preset class activation mapping network model, the method further comprises: If the positive sample image is a car and a truck, the training sample image set is input into the first type of activation mapping network sub-model for training, and a first prediction sub-model is generated when the first type of activation mapping network sub-model converges. If the positive sample image is a car and a car, the training sample image set is input into the second type of activation mapping network sub-model for training, and a first prediction sub-model is generated when the second type of activation mapping network sub-model converges. If the positive sample image is a truck and a truck, the training sample image set is input into the third type of activation mapping network sub-model for training, and a first prediction sub-model is generated when the third type of activation mapping network sub-model converges.

5. The traffic accident prediction method based on artificial intelligence according to claim 3, wherein if the vehicle type label is the first label, the accident vehicle image is input into the first prediction sub-model for prediction to obtain the prediction result, comprising: extracting a saliency map and a corresponding feature map of the accident vehicle image using the first prediction sub-model; determining the prediction result based on the saliency map and the feature map.

6. The traffic accident prediction method based on artificial intelligence according to claim 1, wherein after the step of calling the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image to obtain a prediction result, further comprising: determining an accident level corresponding to the accident vehicle image according to the prediction result; determining an accident vehicle claim strategy according to a preset accident level and claim strategy mapping relationship table and the accident level. comprising: an acquisition module configured to acquire an accident vehicle image; 7.A traffic accident prediction device based on artificial intelligence, characterized by, an identification module configured to call a vehicle type identification sub-model in a trained traffic accident prediction model to identify the accident vehicle image to obtain a vehicle type label, wherein the traffic accident prediction model comprises a vehicle type identification sub-model and a plurality of prediction sub-models, one vehicle type label corresponds to one prediction sub-model, and the prediction sub-model is trained according to a class activation mapping network model framework, wherein the prediction sub-model is obtained by training a class activation latent mapping network (CALM) model; ​ ​ The prediction module is configured to call the prediction sub-model corresponding to the vehicle type label to predict the accident vehicle image, the vehicle type label comprises a first label, a second label or a third label, the first label is a motor vehicle pair, the second label is a motor vehicle and non-motor vehicle pair, and the third label is a non-motor vehicle pair, the motor vehicle comprises a car or a truck, and the non-motor vehicle comprises an electric vehicle or a bicycle; the prediction sub-model comprises a first prediction sub-model corresponding to the first label, a second prediction sub-model corresponding to the second label or a third prediction sub-model corresponding to the third label; if the vehicle type label is the first label, the accident vehicle image is input into the first prediction sub-model for prediction to obtain a prediction result; if the vehicle type label is the second label, the accident vehicle image is input into the second prediction sub-model for prediction to obtain the prediction result; and if the vehicle type label is the third label, the accident vehicle image is input into the third prediction sub-model for traffic accident prediction to obtain the prediction result. 8.A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and characterized in that, the processor, when executing the computer program, implements the steps of the traffic accident prediction method based on artificial intelligence according to any one of claims 1 to 6. 9.A computer readable storage medium, storing a computer program, and characterized in that, the computer program, when executed by a processor, implements the steps of the traffic accident prediction method based on artificial intelligence according to any one of claims 1 to 6.

Images