CN114863680B

CN114863680B - Prediction processing method, prediction processing device, computer equipment and storage medium

Info

Publication number: CN114863680B
Application number: CN202210457213.1A
Authority: CN
Inventors: 钟子宏
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-04-27
Filing date: 2022-04-27
Publication date: 2023-04-18
Anticipated expiration: 2042-04-27
Also published as: CN114863680A

Abstract

The embodiment of the invention discloses a prediction processing method, a prediction processing device, computer equipment and a storage medium, which can be applied to the field of traffic or automatic driving, wherein the method comprises the following steps: acquiring a target data sequence; calling a recurrent neural network model corresponding to the target vehicle in the federal model to perform recurrent prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle; the federal model comprises a recurrent neural network model corresponding to the associated equipment of the target vehicle; obtaining a second accident prediction result aiming at the target vehicle from the associated equipment; the second accident prediction result is obtained by acquiring a reference data sequence by the associated equipment and performing cyclic prediction processing on the reference data sequence by calling a cyclic neural network model corresponding to the associated equipment; and performing combined prediction processing by adopting the first accident prediction result and the second accident prediction result of the target vehicle to obtain a target accident prediction result of the target vehicle, so that the prediction accuracy can be improved.

Description

Prediction processing method, prediction processing device, computer equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a prediction processing method and apparatus, a computer device, and a storage medium.

Background

With the continuous and deep development of computer technology, the computer technology is adopted to predict the events which are likely to occur in the future in the vehicle driving process, so that the safety in the vehicle driving process can be greatly improved, and therefore, the prediction processing of the future events in the vehicle driving process is of great practical significance. However, when the prediction processing is performed on the future event in the driving process of the vehicle, the prediction processing is mostly realized on the basis of the sensor of the vehicle to acquire the sensing data of the environment, and the accuracy of the safety prediction on the basis of the data acquired by the sensor is low, so that how to improve the accuracy of the safety prediction becomes the current research hotspot.

Disclosure of Invention

The embodiment of the invention provides a prediction processing method and device, computer equipment and a storage medium, which can improve the accuracy of prediction.

In one aspect, an embodiment of the present invention provides a prediction processing method, including:

acquiring a target data sequence, wherein the target data sequence is composed of vehicle data generated by the target vehicle at different driving moments in the driving process;

calling a recurrent neural network model corresponding to the target vehicle in the federal model to perform recurrent prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle; the federal model comprises a recurrent neural network model corresponding to the associated equipment of the target vehicle;

obtaining a second accident prediction result for the target vehicle from the correlation device; the second accident prediction result is obtained by acquiring a reference data sequence by the associated equipment and performing cyclic prediction processing on the reference data sequence by calling a cyclic neural network model corresponding to the associated equipment;

and performing combined prediction processing by adopting the first accident prediction result and the second accident prediction result of the target vehicle to obtain a target accident prediction result of the target vehicle.

In another aspect, an embodiment of the present invention provides a prediction processing apparatus, including:

an acquisition unit, configured to acquire a target data sequence, where the target data sequence is composed of vehicle data generated by the target vehicle at different driving times during driving;

the processing unit is used for calling a recurrent neural network model corresponding to the target vehicle in the federal model to carry out recurrent prediction processing on the target data sequence so as to obtain a first accident prediction result of the target vehicle; the federal model comprises a recurrent neural network model corresponding to the associated equipment of the target vehicle;

the processing unit is further used for obtaining a second accident prediction result aiming at the target vehicle from the correlation equipment; the second accident prediction result is obtained by acquiring a reference data sequence by the associated equipment and performing cyclic prediction processing on the reference data sequence by calling a cyclic neural network model corresponding to the associated equipment;

the processing unit is further configured to perform joint prediction processing by using the first accident prediction result and the second accident prediction result of the target vehicle to obtain a target accident prediction result of the target vehicle.

In still another aspect, an embodiment of the present invention provides a computer device, including a processor, an input device, an output device, and a memory, where the processor, the input device, the output device, and the memory are connected to each other, where the memory is used to store a computer program that supports the computer device to execute the above method, the computer program includes program instructions, and the processor is configured to call the program instructions to perform the following steps:

acquiring a target data sequence, wherein the target data sequence is formed by vehicle data generated by the target vehicle at different driving moments in the driving process;

In yet another aspect, an embodiment of the present invention provides a computer-readable storage medium, in which program instructions are stored, and when the program instructions are executed by a processor, the program instructions are used to execute the prediction processing method according to the first aspect.

In the embodiment of the application, after a target data sequence is obtained, a cyclic neural network in the target vehicle is called to perform cyclic prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle, and a second accident prediction result aiming at the target vehicle is obtained from associated equipment of the target vehicle, and after the target vehicle obtains the first accident prediction result and the second accident prediction result, joint prediction processing can be performed by combining the first accident prediction result and the second accident prediction result to obtain the target accident prediction result of the target vehicle, so that when the target vehicle performs accident prediction, the accuracy of the prediction result can be improved by combining not only vehicle running characteristics and driving characteristics of an operation object collected by the target vehicle, but also image data of a road camera and sensing data of equipment such as sensing tests and the like, and prediction of single vehicle accidents based on federal learning is realized.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a schematic diagram of a prediction processing system according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart diagram of a prediction processing method according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a prediction processing method according to an embodiment of the present invention;

fig. 4 is a schematic block diagram of a prediction processing apparatus according to an embodiment of the present invention;

FIG. 5 is a schematic block diagram of a computer device provided by an embodiment of the present invention.

Detailed Description

The embodiment of the application provides a prediction processing method, so that a computer device where a target vehicle is located can perform cyclic prediction processing on a target data sequence associated with the target vehicle by calling a cyclic neural network corresponding to the target vehicle in a federal model, so as to obtain a first accident prediction result of the target vehicle, and in addition, the computer device can also obtain a second accident prediction result aiming at the target vehicle from associated devices of the target vehicle, wherein the target data sequence and a reference data sequence used for predicting and obtaining the second accident prediction result are acquired and processed by different devices, so that when the computer device performs accident prediction on the target vehicle, the accuracy of the finally obtained target accident prediction result can be improved by acquiring the first accident prediction result and the second accident prediction result acquired and predicted by the different devices when performing joint prediction processing subsequently by using the first prediction result and the second prediction result. In one embodiment, the prediction processing method may be applied in the prediction processing system shown in fig. 1, the target vehicle may be a vehicle as indicated by 10 in fig. 1, and the associated device of the target vehicle may be any one of the roadside sensors or image capturing devices as indicated by 11 in fig. 1. It can be understood that the prediction processing System shown in fig. 1 is also an Intelligent Transportation System (ITS), which is a comprehensive Transportation System that effectively and comprehensively applies advanced scientific technologies (information technology, computer technology, data communication technology, sensor technology, electronic control technology, automatic control theory, operations research, artificial intelligence, etc.) to Transportation, service control and vehicle manufacturing, and strengthens the connection among vehicles, roads and users, thereby forming a comprehensive Transportation System that ensures safety, improves efficiency, improves environment and saves energy.

In one embodiment, the target vehicle may also be referred to as a computer device, or the processing device in the target vehicle may also be referred to as a computer device, that is, the target vehicle and the computer device mentioned in the embodiments of the present application are not distinguished; in addition, the federal model is a machine learning model for performing model training and model calling based on federal learning, wherein federal learning refers to the combination of distributed machine learning, cryptography, an incentive mechanism based on financial rules and game theory to solve the use problem of scattered data, so that the model for performing training and prediction processing based on federal learning can realize distributed model training and prediction processing, and in addition, the privacy mechanism in federal learning is combined to effectively protect data security, so that the effective training of the distributed federal model is realized on the premise of not actually exchanging data (such as not actually exchanging a target data sequence and a reference data sequence), and the training precision of the federal model is effectively improved on the premise of protecting data security.

In one embodiment, the federal model includes recurrent neural network models respectively deployed in the target vehicle (i.e., computer devices) and recurrent neural network models deployed in associated devices of the target vehicle, where the number of the associated devices of the target vehicle is one or more, and then the recurrent neural network models deployed in the associated devices include recurrent neural network models deployed in the associated devices, for example, if the associated devices of the target vehicle include road-side sensors and image capturing devices, then the federal model includes not only the recurrent neural network model corresponding to the target device, but also the recurrent neural network model corresponding to the road-side sensors and the recurrent neural network model corresponding to the image capturing devices. In one embodiment, the association relationship between the target vehicle and the association device may be determined based on a driving section where the target vehicle is located, wherein all of the road side sensors and the image acquisition devices of the driving section where the target vehicle is located may be used as the association devices of the target vehicle, or alternatively, the road side sensors and the image acquisition devices in the current driving section (i.e., a part of the driving section where the target vehicle is located) of the target vehicle may be used as the association devices of the target vehicle.

In one embodiment, in order to ensure data security of different data in the same device and avoid the problem that the different types of data affect each other in the accident prediction processing process, thereby causing an error in the accident prediction processing, different recurrent neural network models may also be deployed to perform the accident prediction processing on the different types of data sequences, and specifically, if the target data sequence corresponding to the target vehicle is one or more of a driving data sequence and an operation data sequence, the recurrent neural network model of the target vehicle performing the accident prediction processing on the target data sequence may include one or more of a driving prediction model for processing the driving data sequence and an operation prediction model for processing the operation data sequence.

In one embodiment, the Recurrent Neural Network (or the Recurrent Neural Network model) is a Network model implemented based on a Recurrent Neural Network (RNN) algorithm, where the RNN algorithm is a Recurrent Neural Network in which sequence data is input, such as the above target data sequence or reference data sequence, and all nodes (cyclic units) are connected in a chain manner after recursion (recursion) is performed in an evolution direction of the sequence. The RNN algorithm is a cyclic prediction implemented by an activation function and a smoothing (sigmoid) function, where the activation function is an output function defining the current node at a given input or set of inputs, such as a logic function and an arctan function, and the sigmoid function is a common biological sigmoid function, also called a sigmoid growth curve. In information science, due to the nature of single increment and single increment of inverse function, sigmoid function is often used as the activation function of neural network, mapping variable between 0 and 1, as is common: a logistic regression function. The accident prediction mentioned in the embodiment of the present application refers to a traffic accident prediction, where a traffic accident refers to an event that a vehicle generates a certain loss due to an error or an accident on a road, and the occurrence of the traffic accident may be caused by an irregular operation or may be caused by an incorruptable force factor, so that based on the diversity of the traffic accident, when the prediction processing is performed based on the federal model in the embodiment of the present application, different types of data sequences are also combined to ensure the accuracy and reliability of the accident prediction result, where the traffic accident may be, for example, a collision accident, an explosion accident, or an accident due to scratch, and it can be understood that, when the accident prediction processing is performed on different traffic accidents, different required data sequences may be collected and different federal models may be called to perform the prediction processing, and in the embodiment of the present application, the collision accident is mainly taken as an example, and the data sequence required in the collision accident prediction is taken as an example to be described in detail.

In one embodiment, the data sequence is a data queue arranged based on the acquisition time of data, and when the accident prediction processing is performed on the target data sequence by calling a corresponding recurrent neural network model, the corresponding time based on each data in the data sequence is sequentially input into the recurrent neural network model for processing, and the previous data in the data sequence is output from the recurrent neural network model and is used as the prediction input of the next stage of the recurrent neural network model, so that the recurrent neural network model can identify each data in the data sequence and obtain the corresponding accident prediction result after each data in the data sequence is identified.

Referring to fig. 2, a schematic flow chart of a prediction processing method proposed in the embodiment of the present application is shown, where the prediction processing method can be executed by the target vehicle (i.e. a computer device), and the target vehicle will mainly go through eight stages as shown in fig. 3 when executing the prediction processing method: a data input stage, a sample construction stage, a federal learning-based RNN model construction stage, a federal logarithmic loss model construction stage, a federal RNN gradient model construction stage, a federal RNN model training and testing stage, a federal learning-based RNN model prediction stage, and a traffic accident early warning stage, in which, the prediction processing method mentioned in the embodiment of the present application will be described in detail with reference to fig. 2 and 3, as shown in fig. 2, the prediction processing method may specifically include:

s201, acquiring a target data sequence, wherein the target data sequence is formed by vehicle data generated by a target vehicle at different driving moments in the driving process.

In one embodiment, in the data input stage, the target vehicle inputs the target data sequence into the target vehicle, so as to perform accident prediction processing on the target data sequence by calling a recurrent neural network model corresponding to the target vehicle in the federal model, and obtain a first accident prediction result. The target vehicle is involved in inputting a target data sequence in both a model training process for a federal learning model and a model application process (i.e., the traffic accident early warning stage), but the target data sequence input in the model training process and the model application process has a certain difference. First, in the model training process, the target data sequence input to the target vehicle is obtained through cloud database log data of a vehicle end (i.e., the target vehicle), and in one embodiment, the target data sequence of the target vehicle may be obtained from the cloud database log of the vehicle end based on a license plate number or a vehicle unique Identifier (ID) of the target vehicle, and input to the target vehicle for subsequent model training. In a specific implementation, when the target vehicle acquires the target data sequence, the running of the target vehicle can be acquired firstlyData series and operation data series for the target vehicle, in one embodiment, the travel data series input to the target vehicle (assumed by X) ^W Expressed) comprises one or more of: the vehicle's own speed, the speeds Of the front and rear vehicles, the road surface image collected by the vehicle, the distance between the front and rear vehicles, the distance between the left and right vehicles, the Point Of Interest (POI) data Of the vehicle, the length and width Of the vehicle, the weight Of the vehicle, the fuel consumption Of the vehicle, the number Of the vehicles on the vehicle, the number Of the real-load people Of the vehicle, the maximum speed Of the vehicle, the displacement Of the vehicle, the average speed, the speed limit data, the overspeed data, the traffic flow data, etc. and the operation data sequence (assumed to be X) input to the target vehicle ^U For example) may comprise one or more of the following: the method comprises the steps of image data in a vehicle during driving of a vehicle owner, the number of times of clicking a mobile phone (a vehicle end screen) during driving, the number of times of adjusting a seat during driving, the number of times of adjusting an air conditioner, listening to music data, playing video data, stepping on a brake, stepping on an oil, the swing amplitude of a steering wheel, the swing number of the steering wheel, the number of times of shifting operations, and the like.

It is understood that the running data series input to the target vehicle may be used to reflect the running characteristics of the target vehicle, and the operation data series input to the target vehicle may be used to reflect the operation behavior characteristics of the operation object corresponding to the target vehicle. After the target data sequence is input into the target vehicle, the target vehicle can also obtain an accident tag of the target vehicle, wherein the accident tag comprises tag generation time which is used for indicating the moment when the target vehicle has an accident, and the accident tag can be marked as Y _t Wherein Y is _t = 0,1, event label Y _t When the value of (1) is 0, it means that the target vehicle does not have a traffic accident at time t, and the accident flag Y is set _t If the value of (1) is less than or equal to 1, the target vehicle is determined to have a traffic accident at time t.

After the target vehicle acquires the running data sequence of the target vehicle, the operation data sequence aiming at the target vehicle and the accident label of the target vehicle, the target vehicle can generate the accident label and the label contained in the accident label based on the accident label in the sample construction stageThe accident tag is associated with the travel data sequence and the operation data sequence in time, respectively, and it can be understood that the process of associating the travel data sequence and the operation data sequence based on the accident tag, that is, according to the accident tag Y _t And a process of associating the running data sequence with the operation data sequence according to the vehicle identifier of the target vehicle, that is, the generated target data sequence includes the running data sequence associated with the accident tag and the operation data sequence including the accident tag.

In one embodiment, if the travel data sequence is composed of travel data in a time period of 0 to T, the operation data sequence is composed of operation data in a time period of 0 to T, T >0, if the tag generation time of the accident tag is T, and 0-T, then the target vehicle may associate the accident tag with the travel data sequence in a time period of 0 to T as the travel data sequence after the accident tag is associated according to the tag generation time T of the accident tag and as the operation data sequence after the accident tag is associated according to the tag generation time T of the accident tag when the accident tag is associated with the travel data sequence and the operation data sequence included in the accident tag, respectively. For example, if T takes 10 minutes, the travel data sequence and the operation data sequence are travel data and operation data within 0 to 10 minutes, and further, if T takes 8 minutes, the travel data within 0 to 8 minutes and the accident tag may be associated with each other, and the target data sequence may be generated after the operation data within 0 to 8 minutes and the accident tag are associated with each other.

In one embodiment, after the target vehicle associates the accident tag with the driving data sequence and the operation data sequence, the target vehicle may generate the target data sequence according to the driving data sequence after associating the accident tag and the operation data sequence after associating the accident tagAnd performing sequence segmentation processing on the labeled operation data sequence to respectively obtain one or more running sample sequences and one or more operation sample sequences, then taking any one of the one or more running sample sequences and any one of the one or more operation sample sequences as a target data sequence of the target vehicle, and further subsequently adopting the target data sequence to train the federal model. The driving data sequence after the accident tag is associated with and the operation data sequence after the accident tag is associated with can be referred to as vehicle sample data S _t The subsequent sequence segmentation process may be a process of calling a certain proportion to perform random segmentation, and in general, the sequence segmentation process may obtain two parts, one part is a training sample (i.e. a target data sequence) with a proportion of x, and the two parts can be marked as

Another part is a test sample in a ratio of 1-x which can be marked as +>

In the data input stage, data input is also performed on the associated device of the target vehicle, in one embodiment, the associated device related to the target device may be one or both of the image acquisition device (such as a road camera and the like) and the roadside sensor shown in fig. 1, and then the data sequence input to the image acquisition device in the data input stage is an image data sequence which can be denoted as X ^V And may specifically include one or more of the following: the behavior image data of an operation object in a target vehicle and the image data of the target vehicle on a road are acquired by a road camera, the road condition, the data of the lane where the operation object is located, the speed limit data of the lane where the operation object is located, the average lane change frequency of the vehicle within 1 minute, the number of times that a vehicle owner (namely the operation object of the target vehicle) does not wear a safety belt and the like are extracted by a Convolutional Neural Network (CNN) image processing method. In addition, the data sequence input to the roadside sensor at the data input stage is a sensing data sequence, which may be recordedIs X ^C And specifically includes one or more of vehicle speed data of the target vehicle passing through the section, speed limit data of the section, average vehicle speed data of the passing through the section, the number of times of overspeed of the target vehicle, and the like. It will also be appreciated that the image data sequence acquired by the image acquisition device, i.e., the sensed data sequence acquired by the roadside sensor, may be used to reflect the driving environment characteristics of the target vehicle during driving, and to reflect the driving state of the target vehicle during the corresponding driving.

In the data input stage, the image acquisition equipment is respectively enabled to acquire corresponding image data sequences, and after the roadside sensor acquires corresponding sensing data sequences, the image data sequences and the sensing data sequences are also enabled to be according to the accident label Y in the sample construction stage _t The data processed in the sample construction phase comprises a sequence of driving data X ^W Operating on a data sequence X ^U And a sequence of image data X ^V Sensing data sequence X ^C And according to the accident label Y _t For data sequence { X ^W ，X ^U ，X ^V ，X ^C Dividing the correlation and random proportion to obtain a part of training samples with proportion a, which can be marked as

Another part is a test sample with a ratio of 1-a, which can be marked as->

Based on the protection strategy for data privacy security in the federal learning process, when the corresponding data sequence is associated and randomly divided according to the accident label, the processing process of the corresponding data sequence is executed by the device into which the corresponding data sequence is input, such as the driving data sequence X ^W And operation data sequence X ^U Are all input into the target vehicle in the data input phase, then the driving data sequence X is addressed ^W And operation data sequence X ^U Is processed by the target vehicleSo that the target vehicle performs recognition processing on the processed data sequence (i.e., the target data sequence) through the recurrent neural network model corresponding to the target vehicle in the federal model to obtain a first accident prediction result, i.e., the step S202 is performed. Likewise, the image data sequences X are respectively processed at the associated devices of the target vehicles ^V And a sensing data sequence X ^C After the processing, the corresponding associated equipment can also be based on the processing result, and the associated equipment can also call a recurrent neural network model corresponding to the associated equipment in the federal model to identify and process the processed data sequence and obtain a second accident prediction result. After the associated device identifies the second accident prediction result, the target vehicle may proceed to step S203 by obtaining the second accident prediction result from the associated device, so that the target vehicle may implement the image data sequence X ^V And a sensing data sequence X ^C The accuracy of the accident prediction result obtained by final prediction can be improved by indirect utilization of the method.

S202, calling a recurrent neural network model corresponding to the target vehicle in the federal model to carry out recurrent prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle; the federal model contains a recurrent neural network model corresponding to the associated devices of the target vehicle.

S203, acquiring a second accident prediction result aiming at the target vehicle from the associated equipment; and the second accident prediction result is obtained by acquiring a reference data sequence by the associated equipment and performing cyclic prediction processing on the reference data sequence by calling a cyclic neural network model corresponding to the associated equipment.

In step S202 and step S203, after the sample construction phase is completed, a federal model for accident prediction is constructed in an RNN model construction phase of federal learning, it can be understood that the federal model is a model jointly trained by multiple parties, and therefore, the constructed federal model is also a recognition model for multiple parties, in an embodiment, the federal model for accident prediction includes a recurrent neural network model corresponding to a target vehicle and a recurrent neural network model corresponding to associated equipment of the target vehicle, and a specific model expression may be as shown in formula 1:

wherein, the expression corresponding to the recurrent neural network model corresponding to the target vehicle in the federal model comprises

And

the expression shown, and the recurrent neural network model corresponding to the device associated with the target vehicle corresponds to an expression that is &>

And

and P is an expression when joint prediction is carried out based on the first accident prediction result and the second accident prediction result. In one embodiment, is selected>

Representing a sequence of driving data, X, of the target vehicle at the current time t ^U A sequence of operating data representing a target vehicle>

Represents a sequence of image data and->

Represents a sensed data sequence, and W _t Representing the characteristic weight, V, of the sequence of driving data obtained in the iteration t +1 (i.e. the current round) _t Representing the characteristic weight, U, of the image data sequence resulting from the t +1 (i.e. current round) iteration _t Characteristic weight and Ct representation of operation data sequence obtained by t +1 th iteration (namely current round)And (3) obtaining the characteristic weight of the sensing data sequence in the t +1 th (namely the current round) iteration. In addition, the first and second substrates are,

hidden layer weight vector representing a preceding run data sequence, based on the previous run data sequence>

A hidden layer weight vector representing a previous round of image data sequence, based on the value of the hidden layer weight vector->

Hidden layer weight vector representing a sequence of operational data of a previous round, -a>

Hidden layer weight vector representing a preceding round of sensory data sequence, based on the hidden layer weight vector in the previous round of sensory data>

Hidden layer parameter, representing a preceding round of a sequence of driving data, is evaluated>

Hidden layer parameter, representing a preceding round of image data sequence, in conjunction with a video coding unit>

Hidden layer parameter, representing a sequence of operational data of a previous round, is evaluated>

Hidden layer parameters, { Θ, representing the previous round of sensory data sequence _r L r ∈ (W, V, U, C) } denotes the model weight to be trained (i.e., the prediction weight). The target vehicle adopts the privacy mechanism of federal learning and the model construction method in model training, the privacy mechanism of federal learning is effectively utilized to protect the data security of users, and the RNN algorithm model based on federal learning is written and constructed on the premise of no data interaction.

Based on the federal model as shown in formula 1, the federal modelThe recurrent neural network model corresponding to the target vehicle in the model comprises: a traveling prediction model corresponding to the target vehicle, corresponding to equation 1

The corresponding expression, and the corresponding operation prediction model of the target vehicle, corresponding to ^ 1>

And the target data sequence comprises a running sample sequence and an operation sample sequence, so that when the target vehicle calls a recurrent neural network model corresponding to the target vehicle in the federal model to perform cyclic prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle, the running prediction model corresponding to the target vehicle can be called to perform cyclic prediction processing on the running sample sequence to obtain a running accident prediction result of the target vehicle, namely the result of judging whether the target vehicle is in a live condition or not>

Corresponding values are taken; in addition, the target vehicle can call an operation prediction model corresponding to the target vehicle to perform circulation prediction processing on the operation sample sequence to obtain an operation accident prediction result of the target vehicle, namely ^ based on the operation accident prediction result ^ based on the operation sample sequence>

And correspondingly, the target vehicle can take the driving accident prediction result and the operation accident prediction result as the first accident prediction result of the target vehicle.

In one embodiment, the association device of the target vehicle includes one or both of an image acquisition device and a roadside sensor, and the recurrent neural network model corresponding to the association device includes an image prediction model (i.e., the above equation 1)

Corresponding expression) and a sensory prediction model (corresponding model expression is ≥>

Corresponding expression) of the expression; then, the reference data sequence contains one or both of an image data sequence acquired by the image acquisition device and a sensing data sequence acquired by the roadside sensor; therefore, the second accident prediction result obtained from the associated equipment comprises: image Accident prediction result (i.e. </or >) by image prediction model using image data sequence cyclic prediction processing>

Corresponding to the value of the expression) and a sensing accident prediction result (namely, the value is greater than or equal to the preset value) obtained by the sensing prediction model through the cyclic prediction processing of the sensing data sequence>

Values corresponding to expressions).

In one embodiment, the first accident prediction result and the second accident prediction result are obtained by calling a corresponding recurrent neural network model to perform recurrent prediction processing on the corresponding sequence data, wherein the recurrent prediction processing is specifically performed by: feature weights to the corresponding prediction model (i.e., W as described above) _t ，V _t ，U _t ，C _t ) Hidden layer weights (i.e., as described above)

) Hiding the layer parameter (i.e.. Above->

) An initialization process is performed, and in one embodiment, the initialized feature weight may be denoted as [ W ] _t ，V _t ，U _t ，C _t ] ₀ And [ W ] _t ，V _t ，U _t ，C _t )] ₀ ＝[[1,…,1] ^T ,[1,…,1] ^T ,[1,…,1] ^T ,[1,…,1] ^T ]The initialized hidden layer weight can be denoted as [ Z ] ^W ，Z ^V ，Z ^U ，Z ^C ] ₁ And is andsatisfy [ Z ] ^W ，Z ^V ，Z ^U ，Z ^C ] ₁ ＝[[1,…,1] ^T ,[1,…,1] ^T ,[1,…,1] ^T ,[1,…,1] ^T ]The initialized hidden layer parameter is a parameter matrix, which can be recorded as [ H ] ^W ，H ^V ，H ^U ，H ^C ] ₀ And satisfies formula 2.

After the target vehicle initializes the feature weight, the hidden layer weight and the hidden layer parameter, the target vehicle may further adopt the initialized feature weight [ W ] as shown in formula 1 _t ，V _t ，U _t ，C _t ] ₀ And initialized hidden layer weight [ Z ] ^W ，Z ^V ，Z ^U ，Z ^C ] ₁ For the first sample of the sample sequence and the initialized hidden layer parameter [ H ] ^W ，H ^V ，H ^U ，H ^C ] ₀ And performing weighted summation to obtain a first accident prediction result, and further taking the first accident prediction result as a hidden layer parameter of the next round and combining a second sample of the sample sequence to perform cyclic prediction processing.

In one embodiment, after the target vehicle obtains the first accident prediction result and the second accident prediction result, the target vehicle may perform a joint prediction process using the first accident prediction result and the second accident prediction result to obtain a target accident prediction result of the target vehicle, i.e., proceed to step S204.

And S204, performing combined prediction processing by adopting the first accident prediction result and the second accident prediction result of the target vehicle to obtain a target accident prediction result of the target vehicle.

The target accident prediction result obtained by performing combined prediction processing by using the first accident prediction result and the second accident prediction result can be obtained by

Carry out and representAnd->

Satisfy +>

Obtaining a target accident prediction result->

Then, model training can be carried out on the federated model iteratively based on the target accident prediction result, and in specific implementation, the target vehicle can obtain an accident label Y associated with the target vehicle after obtaining the target accident prediction result _t (ii) a And then comparing the accident label associated with the target vehicle with the target accident prediction result of the target vehicle, and then optimizing the federal model by adopting the comparison result until the optimized federal model is obtained. In one embodiment, when the target vehicle optimizes the federal model by using the comparison result, if the comparison result indicates that the accident label is different from the target accident prediction result, the model loss value of the target vehicle can be calculated according to the target data sequence, and a first prediction weight set by the federal model and aiming at the first accident prediction result is updated by using a gradient descent algorithm; and; and sending the comparison result to the associated equipment so that the associated equipment determines an associated model loss value according to the comparison result and the reference data sequence, and updating a second prediction weight set by the federated model and aiming at a second accident prediction result by adopting a gradient descent algorithm and the associated model loss value so as to optimize the federated model.

In one embodiment, the corresponding loss function based on the above-described federal model can be represented by equation 3:

wherein L is _W Loss function representing a running prediction model, L _U Representing loss of operating prediction modelFunction, L _V Loss function, L, representing an image prediction model _C Represents the loss function of the sensing prediction model, and L _WV Then represents the interaction loss function between the driving prediction model and the image prediction model, L _WU Representing the interaction loss function between the driving prediction model and the operation prediction model, L _WC Representing the interaction loss function between the driving prediction model and the sensing prediction model, L _VU Representing the interaction loss function between the image prediction model and the sensing prediction model, L _VC Representing the interaction loss function between the image prediction model and the sensing prediction model, L _UC An interaction loss function between the operational prediction model and the sensory prediction model is represented, and n represents a total number of vehicles. Then, the model loss value of the target vehicle determined by the target vehicle includes: the target vehicle corresponds to a running prediction model (the corresponding model expression is L) _W ) The target vehicle corresponds to the operation prediction model (corresponding model expression is L) _U ) And a mutual loss value between the traveling prediction model and the operation prediction model (corresponding function expression is L) _WU )；

The associated model loss values include: image prediction model (corresponding function expression is L) _V ) The loss value of (2), a sensing prediction model (corresponding function expression is L) _C ) The loss value of (2), the interaction loss value between the image prediction model and the sensing prediction model (the corresponding function expression is L) _VC ) A driving prediction model and an image prediction model (corresponding function expression is L) _WV ) The value of the mutual loss between the driving prediction model and the sensing prediction model (the corresponding function expression is L) _WC ) Value of interaction loss between, operation prediction model and image prediction model (corresponding function expression is L) _VU ) Value of interaction loss therebetween, and operation prediction model and sensing prediction model (corresponding function expression is L) _UC ) The value of the interaction loss between.

In one embodiment, based on equation 1, since the target vehicle includes a driving prediction model and an operation prediction model in the federal model, the corresponding first prediction weight includes: running forecastThe prediction weights of the model and the corresponding prediction weights of the operation prediction model, and the second prediction weight may comprise: the prediction weight of the image prediction model and the prediction weight corresponding to the sensing prediction model; and any prediction weight is used for carrying out weighted summation processing on the accident prediction result of the corresponding prediction model. In one embodiment, Θ can be employed _W,t+1 Prediction weight, theta, representing the current driving prediction model _U,t+1 The prediction weight, Θ, representing the current operating prediction model _V,t+1 The prediction weight, Θ, representing the current image prediction model _C,t+1 And representing the prediction weight of the current sensing prediction model, when updating each prediction weight, combining the regularization parameter λ and the learning rate η, and performing update processing by using a gradient descent algorithm, which can be specifically referred to as equation 4.

Based on the set prediction weight, when the target vehicle performs combined prediction processing by using the first accident prediction result and the second accident prediction result of the target vehicle to obtain the target accident prediction result of the target vehicle, the first prediction weight set for the first accident prediction result and the second prediction weight set for the second accident prediction result in the federal model can be obtained, and then the first prediction weight and the second prediction weight can be used for performing weighted summation processing on the first accident prediction result and the second accident prediction result respectively to obtain the target accident prediction result of the target vehicle, wherein the target accident prediction result is used for indicating the probability of traffic accidents of the target vehicle. In one embodiment, in the model training stage, the prediction weight refers to the prediction weight obtained by the current model training, and after the model training is completed, the prediction weight refers to the prediction weight obtained by the final training

When the target vehicle updates the prediction weights, the prediction weights are initialized firstProcessed, the initialized prediction weight is obtained and is marked as [ theta ] _W ,Θ _V ,Θ _U ,Θ _C ] ₀ The initialized prediction weights also satisfy: [ theta ] A _W ,Θ _V ,Θ _U ,Θ _C ] ₀ ＝[[1,…,1] ^T ,[1,…,1] ^T ,[1,…,1] ^T ,[1,…,1] ^T ]. It can be understood that the process of updating the prediction weight also belongs to the part of performing model training on the federated model, that is, when performing model training on the federated model, after initializing the feature weight, the hidden layer parameter and the prediction weight, the training sample(s) can be input>

And collecting regularization parameters and learning rate to train a federated model, and obtaining a loss value [ L ] of the first iteration _W ,L _V ,L _U ,L _C ,L _WV ,L _WU ,L _WC ,L _VU ,L _VC ,L _UC ] ₁ Then, the update parameter [ theta ] of the prediction weight can be obtained _W ,Θ _V ,Θ _U ,Θ _C ] ₁ So as to obtain the next prediction weight, and repeating the steps until the obtained loss value is the target loss value->

And a final characteristic weight->

The final hidden layer weight->

And can obtain the trained prediction weight based on the cross loss model>

And a final hidden layer parameter>

And then the federal model which is trained can be obtained.

After the target device completes the model training process, the trained federal model can be deployed to a vehicle collision avoidance prediction system, namely a system which is arranged in a vehicle and used for measuring and calculating the collision probability of the vehicle, if the trained federal model can be used for carrying out accident prediction processing on the target vehicle, the target vehicle needs to obtain a corresponding target data sequence, and the trained federal model is called to complete accident prediction aiming at the target vehicle based on the target data sequence. In one embodiment, in the actual application process, when the target vehicle acquires the target data sequence, the identification information of a reference vehicle of the target vehicle may be acquired first, where the reference vehicle of the target vehicle includes one or both of a vehicle traveling before the target vehicle and a vehicle traveling after the target vehicle; then, the target vehicle acquires a driving data sequence and an operation data sequence of the reference vehicle based on the identification information; the travel data series of the reference vehicle and the operation data series of the reference vehicle are set as target data series. In specific implementation, the target vehicle may obtain the license plate numbers of the front and rear vehicles in the driving process through a CNN technology, and then match the driving data sequence and the operation data sequence of the front and rear vehicles from the cloud database based on the license plate numbers as the target data sequence in the actual prediction process, where the target data sequence in this case includes

The travel data and the operational data at the current time are identified.

Target data sequence in acquisition to prediction process

Thereafter, the final feature weights may be employed

The final hidden layer weight->

Trained predictive weight->

And a final hidden layer parameter->

And performing accident prediction processing by the union formula 1 to obtain a target prediction result of the traffic accident of the target vehicle at the current moment, wherein the target prediction result can be a probability vector P { Y } _t+1 And expressing.

In one embodiment, the target accident prediction result is used to indicate a target probability of a traffic accident occurring in the target vehicle, and then the target vehicle may compare the target probability of the traffic accident occurring in the target vehicle indicated by the target accident prediction result with a probability threshold, where the probability threshold may be represented by θ, and the value of the probability threshold θ is generally 0.5 or 0.7, etc., so that the target vehicle may obtain the target probability P { Y } _t+1 After that, when the target probability is greater than or equal to the probability threshold, namely P { Y } _t+1 }>And when theta is reached, outputting traffic prompt information, and reducing the running speed of the target vehicle according to the traffic prompt information, otherwise, the target vehicle considers that the probability of the occurrence of the traffic accident is low, and therefore, no prompt is output. Alternatively, the target vehicle may directly control the target vehicle to reduce the current running speed when the target probability is greater than the probability threshold.

Based on the description of the above prediction processing method embodiment, an embodiment of the present invention also proposes a prediction processing apparatus, which may be a computer program (including program code) running in the above target vehicle, i.e. a computer device. The prediction processing apparatus can be used to execute the prediction processing method shown in fig. 2, please refer to fig. 4, and the prediction processing apparatus includes: an acquisition unit 401 and a processing unit 402.

An obtaining unit 401, configured to obtain a target data sequence, where the target data sequence is composed of vehicle data generated by the target vehicle at different driving moments during a driving process;

the processing unit 402 is configured to invoke a recurrent neural network model corresponding to the target vehicle in a federal model to perform recurrent prediction processing on the target data sequence, so as to obtain a first accident prediction result of the target vehicle; the federal model comprises a recurrent neural network model corresponding to the associated equipment of the target vehicle;

the processing unit 402 is further configured to obtain a second accident prediction result for the target vehicle from the associated device; the second accident prediction result is obtained by acquiring a reference data sequence by the associated equipment and performing cyclic prediction processing on the reference data sequence by calling a cyclic neural network model corresponding to the associated equipment;

the processing unit 402 is further configured to perform joint prediction processing by using the first accident prediction result and the second accident prediction result of the target vehicle, so as to obtain a target accident prediction result of the target vehicle.

In an embodiment, the obtaining unit 401 is specifically configured to:

acquiring a running data sequence of the target vehicle and an operation data sequence aiming at the target vehicle, and acquiring an accident label of the target vehicle, wherein the accident label comprises label generation time which is used for indicating the moment when the target vehicle has an accident;

associating the accident label with the driving data sequence and the operation data sequence respectively based on the accident label and the label generation time contained in the accident label;

and generating the target data sequence according to the driving data sequence associated with the accident label and the operation data sequence associated with the accident label.

In one embodiment, the driving data sequence is composed of driving data in a time period of 0-T, the operation data sequence is composed of operation data in a time period of 0-T, T >0, if the label generation time of the accident label is T, and 0-T; the obtaining unit 401 is specifically configured to:

and according to the tag generation time t of the accident tag, associating the accident tag with the driving data sequence within the time period of 0-t to serve as the driving data sequence after the accident tag is associated, and according to the tag generation time t of the accident tag, associating the accident tag with the operation data sequence within the time period of 0-t to serve as the operation data sequence after the accident tag is associated.

In an embodiment, the obtaining unit 401 is specifically configured to:

respectively carrying out sequence segmentation processing on the driving data sequence associated with the accident label and the operation data sequence associated with the accident label to respectively obtain one or more driving sample sequences and one or more operation sample sequences;

taking any of the one or more sequences of travel samples and any of the one or more sequences of operation samples as a target data sequence for the target vehicle.

In one embodiment, the recurrent neural network model corresponding to the target vehicle in the federal model includes: the target data sequence comprises a running sample sequence and an operation sample sequence; the processing unit 402 is specifically configured to:

calling a running prediction model corresponding to the target vehicle to perform cyclic prediction processing on the running sample sequence to obtain a running accident prediction result of the target vehicle;

calling an operation prediction model corresponding to the target vehicle to perform cyclic prediction processing on the operation sample sequence to obtain an operation accident prediction result of the target vehicle;

and taking the driving accident prediction result and the operation accident prediction result as a first accident prediction result of the target vehicle.

In one embodiment, the association device of the target vehicle comprises one or two of an image acquisition device and a roadside sensor, and the recurrent neural network model corresponding to the association device comprises one or two of an image prediction model and a sensing prediction model;

the reference data sequence comprises one or both of an image data sequence acquired by the image acquisition device and a sensing data sequence acquired by the roadside sensor;

the second accident prediction result obtained from the associated device comprises: and one or two of an image accident prediction result obtained by the image prediction model through the image data sequence cyclic prediction processing and a sensing accident prediction result obtained by the sensing prediction model through the sensing data sequence cyclic prediction processing.

In an embodiment, the processing unit 402 is specifically configured to:

initializing the feature weight, hidden layer weight and hidden layer parameter of the corresponding prediction model;

carrying out weighted summation processing on a first sample of the sample sequence and initialized hidden layer parameters by adopting the initialized characteristic weight and the initialized hidden layer weight to obtain a first accident prediction result;

and taking the first accident prediction result as a hidden layer parameter of the next round, and combining a second sample of the sample sequence to perform cyclic prediction processing.

In an embodiment, the processing unit 402 is specifically configured to:

acquiring a first prediction weight set for the first accident prediction result and a second prediction weight set for the second accident prediction result in the federal model;

and respectively carrying out weighted summation processing on the first accident prediction result and the second accident prediction result by adopting the first prediction weight and the second prediction weight to obtain a target accident prediction result of the target vehicle, wherein the target accident prediction result is used for indicating the probability of the traffic accident of the target vehicle.

In one embodiment, the obtaining unit 401 is further configured to obtain an accident tag associated with the target vehicle;

the processing unit 402 is further configured to compare the accident label associated with the target vehicle with the target accident prediction result of the target vehicle;

the processing unit 402 is further configured to optimize the federated model by using the comparison result until an optimized federated model is obtained.

In an embodiment, the processing unit 402 is specifically configured to:

if the comparison result indicates that the accident label is different from the target accident prediction result, calculating to obtain a model loss value of the target vehicle according to the target data sequence, and updating a first prediction weight which is set by the federal model and aims at the first accident prediction result by adopting a gradient descent algorithm; and;

and sending the comparison result to the associated equipment so that the associated equipment determines an associated model loss value according to the comparison result and the reference data sequence, and updating a second prediction weight, which is set by the federal model and aims at the second accident prediction result, by adopting a gradient descent algorithm and the associated model loss value so as to optimize the federal model.

In one embodiment, the model loss value of the target vehicle comprises: the loss value of the target vehicle corresponding to a running prediction model, the loss value of the target vehicle corresponding to an operation prediction model, and the interaction loss value between the running prediction model and the operation prediction model;

the associated model loss values comprise: a loss value of an image prediction model, a loss value of the sensing prediction model, an interaction loss value between the image prediction model and the sensing prediction model, an interaction loss value between the driving prediction model and the image prediction model, an interaction loss value between the driving prediction model and the sensing prediction model, an interaction loss value between the operating prediction model and the image prediction model, and an interaction loss value between the operating prediction model and the sensing prediction model.

In one embodiment, the first prediction weight comprises: the prediction weight of the running prediction model and the prediction weight corresponding to the operation prediction model;

the second prediction weights include: the prediction weight of the image prediction model and the prediction weight corresponding to the sensing prediction model;

and any prediction weight is used for carrying out weighted summation processing on the accident prediction result of the corresponding prediction model.

In an embodiment, the obtaining unit 401 is specifically configured to:

acquiring identification information of a reference vehicle of the target vehicle, wherein the reference vehicle of the target vehicle comprises one or two of a vehicle running before the target vehicle and a vehicle running behind the target vehicle;

acquiring a running data sequence and an operation data sequence of the reference vehicle based on the identification information;

and taking the running data sequence of the reference vehicle and the operation data sequence of the reference vehicle as target data sequences.

In one embodiment, the target accident prediction result is used to indicate a target probability of a traffic accident occurring for the target vehicle; the processing unit 402 is further configured to compare a target probability that the target accident prediction result indicates that the target vehicle has a traffic accident with a probability threshold;

the processing unit 402 is further configured to output traffic prompt information when the target probability is greater than or equal to the probability threshold, and reduce the driving speed of the target vehicle according to the traffic prompt information.

In this embodiment, after the obtaining unit 401 obtains the target data sequence, the processing unit 402 may invoke a recurrent neural network in the target vehicle to perform recurrent prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle, and obtain a second accident prediction result for the target vehicle from a device associated with the target vehicle, and after the processing unit 402 obtains the first accident prediction result and the second accident prediction result, the processing unit may perform joint prediction processing by combining the first accident prediction result and the second accident prediction result to obtain the target accident prediction result of the target vehicle, so that when performing accident prediction, the accuracy of the prediction result may be improved by combining not only the driving characteristics of the vehicle collected by the target vehicle and the driving characteristics of an operating object, but also the image data of a road camera and the sensing data of a device such as a sensing test, and the like, and the prediction of a single vehicle accident based on federal learning may be implemented, and the federal learning based on the learning mechanism and model building manner may effectively utilize the federal learning mechanism to protect the data security of a user, and implement the security of private data without using other privacy data on the premise of privacy data interaction.

Fig. 5 is a schematic block diagram of a computer device according to an embodiment of the present invention, where the computer device is the target vehicle. The computer device in the present embodiment as shown in fig. 5 may include: one or more processors 501; one or more input devices 502, one or more output devices 503, and memory 504. The processor 501, the input device 502, the output device 503, and the memory 504 are connected by a bus 505. The memory 504 is used for storing a computer program comprising program instructions, and the processor 501 is used for executing the program instructions stored by the memory 504.

The memory 504 may include volatile memory (volatile memory), such as random-access memory (RAM); the memory 504 may also include a non-volatile memory (non-volatile memory), such as a flash memory (flash memory), a solid-state drive (SSD), etc.; the memory 504 may also comprise a combination of the above-described types of memory.

The processor 501 may be a Central Processing Unit (CPU). The processor 501 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a Programmable Logic Device (PLD), or the like. The PLD may be a field-programmable gate array (FPGA), a General Array Logic (GAL), or the like. The processor 501 may also be a combination of the above structures.

In the embodiment of the present invention, the memory 504 is used for storing a computer program, the computer program includes program instructions, and the processor 501 is used for executing the program instructions stored in the memory 504, so as to implement the steps of the corresponding method as described above in fig. 2.

In one embodiment, the processor 501 is configured to call the program instructions to perform:

obtaining a second accident prediction result for the target vehicle from the associated device; the second accident prediction result is obtained by acquiring a reference data sequence by the associated equipment and performing cyclic prediction processing on the reference data sequence by calling a cyclic neural network model corresponding to the associated equipment;

In one embodiment, the driving data sequence is composed of driving data in a time period of 0-T, the operation data sequence is composed of operation data in a time period of 0-T, T >0, if the label generation time of the accident label is T, and 0-T; the processor 501 is configured to call the program instructions for performing:

respectively carrying out sequence segmentation processing on the driving data sequence associated with the accident label and carrying out sequence segmentation processing on the operation data sequence associated with the accident label to respectively obtain one or more driving sample sequences and one or more operation sample sequences;

In one embodiment, the recurrent neural network model corresponding to the target vehicle in the federal model includes: the target data sequence comprises a running sample sequence and an operation sample sequence; the processor 501 is configured to call the program instructions for performing:

obtaining an accident label associated with the target vehicle;

comparing the accident label associated with the target vehicle with a target accident prediction result of the target vehicle;

and optimizing the federal model by adopting a comparison result until the optimized federal model is obtained.

and sending the comparison result to the associated equipment so that the associated equipment determines an associated model loss value according to the comparison result and the reference data sequence, and updating a second prediction weight set by the federal model for the second accident prediction result by adopting a gradient descent algorithm and the associated model loss value so as to optimize the federal model.

In one embodiment, the target accident prediction result is used to indicate a target probability of a traffic accident occurring for the target vehicle; the processor 501 is configured to call the program instructions for performing:

comparing the target probability that the target accident prediction result indicates that the target vehicle has a traffic accident with a probability threshold;

and when the target probability is greater than or equal to the probability threshold, outputting traffic prompt information, and reducing the running speed of the target vehicle according to the traffic prompt information.

Embodiments of the present invention provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method embodiment as shown in fig. 2. The computer-readable storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), or the like.

While the invention has been described with reference to a particular embodiment, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A prediction processing method, comprising:

acquiring a target data sequence, wherein the target data sequence is composed of vehicle data generated by a target vehicle at different driving moments in a driving process;

2. The method of claim 1, wherein said obtaining a target data sequence comprises:

3. The method according to claim 2, wherein the travel data sequence is composed of travel data in a period of 0 to T, the operation data sequence is composed of operation data in a period of 0 to T, T >0, if the tag generation time of the accident tag is T, and 0-T-s; associating the accident label with the driving data sequence and the operation data sequence respectively based on the accident label and the label generation time contained in the accident label, including:

4. The method of claim 2, wherein generating the target data sequence for the target vehicle based on the travel data sequence associated with the accident tag and the operational data sequence associated with the accident tag comprises:

taking any of the one or more travel sample sequences and any of the one or more operation sample sequences as a target data sequence of the target vehicle.

5. The method of claim 1, wherein the target vehicle's corresponding recurrent neural network model in the federal model comprises: the target data sequence comprises a running sample sequence and an operation sample sequence; the calling of the recurrent neural network model corresponding to the target vehicle in the federal model to carry out recurrent prediction processing on the target data sequence to obtain a first accident prediction result of the target vehicle comprises the following steps:

6. The method of claim 1, wherein the associated device of the target vehicle comprises one or both of an image acquisition device and a roadside sensor, and the recurrent neural network model corresponding to the associated device comprises one or both of an image prediction model and a sensing prediction model;

7. The method of claim 5 or 6, wherein the cyclic prediction processing of the sequence of samples using the corresponding prediction model comprises:

initializing the characteristic weight, hidden layer weight and hidden layer parameter of the corresponding prediction model;

8. The method according to claim 1, wherein the performing a joint prediction process using the first accident prediction result and the second accident prediction result of the target vehicle to obtain the target accident prediction result of the target vehicle comprises:

9. The method of claim 1, wherein the method further comprises:

obtaining an accident label associated with the target vehicle;

10. The method of claim 9, wherein the using the comparison results to optimize the federal model includes:

11. The method of claim 10, wherein the model loss value for the target vehicle comprises: the loss value of the target vehicle corresponding to a running prediction model, the loss value of the target vehicle corresponding to an operation prediction model, and the interaction loss value between the running prediction model and the operation prediction model;

the associated model loss values include: a loss value of an image prediction model, a loss value of a sensing prediction model, an interaction loss value between the image prediction model and the sensing prediction model, an interaction loss value between the driving prediction model and the image prediction model, an interaction loss value between the driving prediction model and the sensing prediction model, an interaction loss value between the operating prediction model and the image prediction model, and an interaction loss value between the operating prediction model and the sensing prediction model.

12. The method of claim 10, wherein the first prediction weight comprises: the prediction weight of the driving prediction model and the prediction weight corresponding to the operation prediction model;

13. The method of claim 1, wherein said obtaining a target data sequence comprises:

acquiring identification information of a reference vehicle of the target vehicle, wherein the reference vehicle of the target vehicle comprises one or two of a vehicle running before the target vehicle and a vehicle running after the target vehicle;

14. The method of claim 1, wherein the target accident prediction result is used to indicate a target probability of a traffic accident occurring for the target vehicle; the method further comprises the following steps:

15. A prediction processing apparatus, comprising:

an acquisition unit configured to acquire a target data sequence composed of vehicle data generated by a target vehicle at different travel times during a travel process;

16. A computer device comprising a processor, an input device, an output device and a memory, the processor, the input device, the output device and the memory being interconnected, wherein the memory is configured to store a computer program comprising program instructions, the processor being configured to invoke the program instructions to perform the method of any of claims 1 to 14.

17. A computer-readable storage medium, characterized in that it stores a computer program comprising program instructions which, when executed by a processor, cause the processor to carry out the method according to any one of claims 1 to 14.