Traffic accident detection model optimization method and device, equipment and medium

Abstract

The invention discloses a traffic accident detection model optimization method and device, equipment and a medium, and belongs to the technical field of traffic accident detection.The method comprises the steps that user historical data within a preset period of time and missing detection data which is not correctly recognized as an accident in the user historical data are obtained; establishing a first scoring model according to user historical data, and estimating a time sequence abnormal score of each piece of missing detection data according to the first scoring model to obtain a first score; acquiring historical slight accident data; establishing a second scoring model according to the historical slight accident data, and estimating a time sequence abnormal score of each piece of missing detection data according to the second scoring model to obtain a second score; calculating a weighted score of the first score and the second score to obtain a third score; and sorting the third scores, selecting driving data contained in the TopN data as a new training sample, and training the accident detection model to obtain an optimized accident detection model. According to the invention, acceleration data can be effectively analyzed, and accident missing detection can be identified.

Description

Technical Field

[0001] The present application belongs to the technical field of traffic accident detection, and in particular, relates to a traffic accident detection model optimization method, device, equipment and medium. Background Art

[0002] With the surge in the number of vehicles in recent years, the frequency of traffic accidents has also increased. In order to improve road safety and reduce the occurrence of traffic accidents, various accident detection systems are widely used in traffic monitoring and early warning. These systems usually rely on technologies such as cameras, radars, and geomagnetic sensors to detect and report accidents. However, these technologies have certain limitations in practical applications. For example, cameras are easily affected by weather and light conditions, and radars and geomagnetic sensors may not be able to accurately judge the severity of an accident.

[0003] An accelerometer is a sensor that can measure the acceleration of an object in the forward, turning, and vertical directions. It is widely used in devices such as smartphones and cars. In the field of traffic accident detection, accelerometers can be used to detect abnormal vehicle movements, such as sudden braking and collisions, thereby providing valuable information about the accident.

[0004] However, existing accident detection methods based on accelerometers (such as the Chinese invention patent with publication number CN116363769A) often have missed detections in accident detection. Summary of the invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a traffic accident detection model optimization method, device, equipment and medium, which can obtain and analyze acceleration data, combine historical minor accident data and the user's historical driving data, and effectively identify and score potential missed accidents to facilitate further investigation and analysis, and provide a sample basis for the construction of models for minor accidents and atypical accidents.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] A traffic accident detection model optimization method, the method comprising:

[0008] Acquire user historical data within a preset period of time and missed detection data in the user historical data that are not correctly identified as accidents;

[0009] Establishing an anomaly scoring model with the user historical data to obtain a first scoring model, and estimating a time series anomaly score of each of the missed data with the first scoring model to obtain a first score;

[0010] Obtain historical minor accident data;

[0011] Establishing an abnormality scoring model with the historical minor accident data to obtain a second scoring model, and estimating the time series abnormality score of each missed detection data with the second scoring model to obtain a second score;

[0012] Calculating a weighted score of the first score and the second score with a preset weight to obtain a third score;

[0013] The third score is sorted, and driving data included in a preset number of missed detection data with scores sorted from high to low are selected as new training samples, and the accident detection model is trained to obtain an optimized accident detection model.

[0014] Furthermore, the method comprises:

[0015] Obtaining an analysis time period of each piece of the missed detection data from the missed detection data, and obtaining a three-axis acceleration time series data block of the analysis time period of the missed detection data from the analysis time period of the missed detection data to obtain a first time series data block;

[0016] estimating each first time series data block using the first scoring model to obtain a first score;

[0017] Each first time series data block is estimated using the second scoring model to obtain a second score.

[0018] Furthermore, the method for establishing the first scoring model specifically includes:

[0019] Obtain the user's three-axis acceleration time series data blocks within a recent preset period of time as training samples;

[0020] Cut each three-axis acceleration time series data block into several parts in chronological order;

[0021] Calculate the first information entropy and the first range for each piece of data, and then calculate the second information entropy and the second range for the data rotated to a standard posture, where the standard posture is the posture of the device in a static state with the x and y axes being 0 and the z axis being 1g, where g is the acceleration of gravity;

[0022] Calculate the statistical indicator characteristics of each axis obtained;

[0023] Calculate the first fluctuation intensity of each axis as a whole, and then calculate the second fluctuation intensity of each axis as a whole for the data rotated to the standard posture;

[0024] The first information entropy, the first range, the second information entropy, the second range, the statistical indicator feature, the first fluctuation severity, and the second fluctuation severity are combined, and a normalization operation is performed;

[0025] The normalized features are used as training data, and the autoencoder structure is used to encode and decode the data. The first scoring model of the corresponding user is obtained by optimizing the loss, and the loss function is the mean square error loss function.

[0026] Furthermore, the method for establishing the second scoring model specifically includes:

[0027] Obtain the three-axis acceleration time series data blocks of all historical minor accident data as training samples;

[0028] Cut each three-axis acceleration time series data block into several parts in chronological order;

[0029] Calculate the third information entropy and the third range for each piece of data, and then calculate the fourth information entropy and the fourth range for the data rotated to a standard posture, where the standard posture is a device posture with x and y axes at 0 and z axis at 1g in a static state, where g is the acceleration of gravity;

[0030] Calculate the statistical indicator characteristics of each axis obtained;

[0031] Calculate the third fluctuation intensity of each axis as a whole, and then calculate the fourth fluctuation intensity of each axis as a whole for the data rotated to the standard posture;

[0032] The third information entropy, the third range, the fourth information entropy, the fourth range, the statistical indicator feature, the third fluctuation severity and the fourth fluctuation severity are combined and normalized;

[0033] The normalized features are used as training data, and the autoencoder structure is used to encode and decode the data. The second scoring model of the corresponding user is obtained by optimizing the loss, and the loss function is the mean square error loss function.

[0034] Furthermore, the statistical indicator characteristics include standard deviation, 10th quantile, 25th quantile, 50th quantile, 75th quantile, 90th quantile and mean.

[0035] Furthermore, the preset weight is obtained by quantifying the analysis capabilities of historical missed detections of the first score and the second score.

[0036] Furthermore, the quantification of the analysis capability of the historical missed detection of the first score and the second score specifically includes:

[0037] w1=l1 / (l1+l2), w2=l2 / (l1+l2), where w1 is the weight of the first score, w2 is the weight of the second score, l1 is the number of successful analyses of historical missed accidents using the first score, and l2 is the number of successful analyses of historical missed accidents using the second score.

[0038] On the other hand, the present invention also provides a traffic accident detection model optimization device, the device comprising:

[0039] A first information acquisition module, which acquires user history data within a preset period of time and missed detection data in the user history data that is not correctly identified as an accident;

[0040] A first anomaly scoring module, wherein the first anomaly scoring module establishes an anomaly scoring model with the user historical data to obtain a first scoring model, and estimates a time series anomaly score of each of the missed data with the first scoring model to obtain a first score;

[0041] A second information acquisition module, the second information acquisition module historical minor accident data;

[0042] A second anomaly scoring module, wherein the second anomaly scoring module establishes an anomaly scoring model based on the historical minor accident data to obtain a second scoring model, and estimates a time series anomaly score of each missed detection data based on the second scoring model to obtain a second score;

[0043] A third anomaly scoring module, wherein the third anomaly scoring module calculates a weighted score of the first score and the second score with a preset weight to obtain a third score;

[0044] A model optimization module, wherein the model optimization module sorts the third scores, selects the driving data contained in a preset number of missed detection data with scores sorted from high to low as new training samples, trains the accident detection model, and obtains an optimized accident detection model.

[0045] On the other hand, the present invention also provides a computer device, which includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement any one of the above-mentioned traffic accident detection model optimization methods.

[0046] On the other hand, the present invention also provides a computer-readable storage medium, in which a computer program is stored. The computer program is loaded and executed by a processor to implement any one of the above-mentioned traffic accident detection model optimization methods.

[0047] The beneficial effects of this application are:

[0048] (1) This application calculates information entropy, range and other indicators for the three-axis time series blocks, and calculates the anomaly scores respectively in combination with the user's driving data and the historical minor accident data distribution. The two anomaly scores are then merged to obtain the final anomaly score. The data with a high final anomaly score is used as a new data sample to optimize the traditional accident detection model, thereby more accurately identifying possible missed accidents.

[0049] (2) This application can quickly locate the time point of missed accidents by obtaining the nearby time driving trajectory, driving video analysis and comparison of the Top N data blocks with abnormal scores, improve the efficiency of missed accident analysis, and provide detailed information and evidence for accident investigation and handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flowchart of the traffic accident detection model optimization method according to an embodiment of the present application;

[0051] Figure 2 It is a structural block diagram of the traffic accident detection model optimization device according to an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0053] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0054] Existing accelerometer accident detection methods often miss detections, especially in minor or atypical accident scenarios. These technologies usually fail to conduct in-depth analysis of missed accidents and therefore fail to provide sufficient information to assist in accident investigation and handling.

[0055] In order to solve the above technical problems, the following embodiments of the traffic accident detection model optimization method, device, equipment and medium of the present application are proposed.

[0056] Example 1

[0057] Reference Figure 1 ,like Figure 1 The figure is a flowchart of a method for optimizing a traffic accident detection model according to an embodiment of the present application. The method specifically comprises the following steps:

[0058] S1. Extract missed detection data that have not been correctly identified as accidents within a period of time from the system in batches; missed detection accidents are usually reported by car owners, auto 4S stores, auto repair shops, insurance companies, etc. and stored in the missed detection data table, which is recorded as Table A here. The time range obtained from Table A can be freely defined, such as the past 1 month, the past 3 months, etc.

[0059] S2. For each missed detection data, the analysis time period of the missed detection accident is obtained according to its information feedback time. Generally, it is from 30 days to the information feedback time, and the analysis time period of each missed detection accident is different.

[0060] S3. For each missed accident, a list of three-axis acceleration time series data blocks within the analysis period obtained in S2 is obtained, wherein the data in a single data block includes time and three-axis acceleration information.

[0061] S4. Estimate the anomaly score score1 of each three-axis time series block according to the anomaly scoring model based on user history.

[0062] As an implementation method, in this embodiment, the training of the abnormal scoring model based on user history can be divided into the following steps.

[0063] S401. Obtain the user's historical three-axis acceleration time series data blocks for the past six months as training samples.

[0064] S402. Cut each three-axis timing block into M parts in time sequence. In this example, M is set to 10.

[0065] S403. Calculate the information entropy of each axis of each copy of data and record it as

[0066] x_ent1,…,x_ent10,y_ent1,…,y_ent10,z_ent1,…z_ent10, calculate the range index in the same way, and then calculate the information entropy and range of the data rotated to the standard posture in the same way. The standard posture is the device posture with x and y axes at 0 and z axis at 1g in a static state (g is the acceleration of gravity, about 9.80m / s 2 ). A total of 120 indicators can be obtained in this step.

[0067] S404. Calculate the standard deviation, 10th quantile, 25th quantile, 50th quantile, 75th quantile, 90th quantile, and mean indicators for the 10 indicators of each axis obtained above. A total of 84 indicators can be obtained in this step.

[0068] S405. Calculate the degree of fluctuation for each axis as a whole, and then calculate the same index for the data rotated to the standard posture. A total of 6 indexes can be obtained in this step.

[0069] S406. Merge the features of S403 to S405, a total of 210 dimensions, and perform normalization. In this embodiment, the normalization method uses mean-variance normalization (Z-score Normalization), also known as standardization, to normalize all data to a distribution with a mean of 0 and a variance of 1. That is, it is ensured that the final data has a mean of 0 and a variance of 1.

[0070] S407. The normalized 210-dimensional features are used as training data, and the data is encoded and decoded using the autoencoder structure. In this example, the number of layers of the autoencoder model is set to 5, and the number of neurons is 210, 64, 16, 64, and 210 respectively. The abnormality scoring model of the user is obtained by optimizing the loss. The loss function uses MSE:

[0071]

[0072] where x i represents the input features, Represents the reconstructed output of the autoencoder, and n represents the number of samples.

[0073] S408. Execute S402 to S406 for the data block to be evaluated, and input the result into the anomaly scoring model to obtain the anomaly score score1 of each three-axis sequence block.

[0074] S5. Combined with historical minor accident data or atypical accident data, the abnormal score score2 of each three-axis time series block is obtained using a method similar to S4. The difference from S4 is that the three-axis acceleration time series data blocks of historical minor accidents of all users are used as training samples.

[0075] S6. Combine the two anomaly scores to get the final anomaly score:

[0076] score=w1*score1+w2*score2

[0077] Among them, w1 and w2 are the weights of the two anomaly scores respectively, and the analysis ability of historical missed detections is quantified by this score. w1=l1 / (l1+l2), where l1 is the number of successful analyses of historical missed detection accidents through score1, and l2 is the number of successful analyses of historical missed detection accidents through score2. w2 can be obtained in a similar way.

[0078] S7. Sort the anomaly scores to obtain the top N data blocks with anomaly scores, where N=10 is generally taken.

[0079] S8. Obtain the driving trajectory and driving video of the time near the top 10 data, and use the driving trajectory and driving video as new training samples to optimize the existing accident detection model.

[0080] The optimized accident detection model can detect minor accidents or atypical accidents more effectively.

[0081] This embodiment calculates information entropy, range and other indicators for the three-axis time series blocks, and calculates the anomaly scores respectively in combination with the user driving data and the historical minor accident data distribution. The two anomaly scores are then merged to obtain the final anomaly score. The traditional accident detection model is then optimized with the features required for event detection corresponding to the data with a high final anomaly score, thereby more accurately identifying possible missed accidents.

[0082] Example 2

[0083] Reference Figure 2 ,like Figure 2 The figure is a structural block diagram of a traffic accident detection model optimization device according to an embodiment of the present application, and the device specifically includes the following structures:

[0084] A first information acquisition module, the first information acquisition module acquires user historical data within a preset period of time and missed detection data in the user historical data that is not correctly identified as an accident;

[0085] A first anomaly scoring module, wherein the first anomaly scoring module establishes an anomaly scoring model based on the user's historical data to obtain a first scoring model, and estimates the time series anomaly score of each missed data using the first scoring model to obtain a first score;

[0086] A second information acquisition module, which includes historical minor accident data;

[0087] A second anomaly scoring module, which establishes an anomaly scoring model based on historical minor accident data to obtain a second scoring model, and estimates the time series anomaly score of each missed detection data using the second scoring model to obtain a second score;

[0088] A third anomaly scoring module, which calculates a weighted score of the first score and the second score with a preset weight to obtain a third score;

[0089] The model optimization module sorts the third scores, selects the driving data contained in the preset number of missed detection data with scores sorted from high to low as new training samples, trains the accident detection model, and obtains the optimized accident detection model.

[0090] Example 3

[0091] This preferred embodiment provides a computer device, which can implement the steps in any embodiment of the traffic accident detection model optimization method provided in the embodiments of the present application. Therefore, the beneficial effects of the traffic accident detection model optimization method provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0092] Example 4

[0093] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by controlling related hardware through instructions, and the instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. To this end, an embodiment of the present invention provides a storage medium, which stores a plurality of instructions, and the instructions can be loaded by a processor to execute the steps of any embodiment of the traffic accident detection model optimization method provided in the embodiment of the present invention.

[0094] The storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0095] Since the instructions stored in the storage medium can execute the steps in any traffic accident detection model optimization method embodiment provided in the embodiments of the present invention, the beneficial effects that can be achieved by any traffic accident detection model optimization method provided in the embodiments of the present invention can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0096] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A traffic accident detection model optimization method, characterized in that: The method comprises: Acquire user historical data within a preset period of time and missed detection data in the user historical data that are not correctly identified as accidents; Establishing an anomaly scoring model with the user historical data to obtain a first scoring model, and estimating a time series anomaly score of each of the missed data with the first scoring model to obtain a first score; Obtain historical minor accident data; Establishing an abnormality scoring model with the historical minor accident data to obtain a second scoring model, and estimating the time series abnormality score of each missed detection data with the second scoring model to obtain a second score; Calculating a weighted score of the first score and the second score with a preset weight to obtain a third score; The third score is sorted, and driving data included in a preset number of missed detection data with scores sorted from high to low are selected as new training samples, and the accident detection model is trained to obtain an optimized accident detection model.

2. The traffic accident detection model optimization method according to claim 1, characterized in that: The method comprises: Obtaining an analysis time period of each piece of the missed detection data from the missed detection data, and obtaining a three-axis acceleration time series data block of the analysis time period of the missed detection data from the analysis time period of the missed detection data to obtain a first time series data block; estimating each first time series data block using the first scoring model to obtain a first score; Each first time series data block is estimated using the second scoring model to obtain a second score.

3. The traffic accident detection model optimization method according to claim 2, characterized in that: The method for establishing the first scoring model specifically includes: Obtain the user's three-axis acceleration time series data blocks within a recent preset period of time as training samples; Cut each three-axis acceleration time series data block into several parts in chronological order; Calculate the first information entropy and the first range for each piece of data, and then calculate the second information entropy and the second range for the data rotated to a standard posture, where the standard posture is the posture of the device in a static state with the x and y axes being 0 and the z axis being 1g, where g is the acceleration of gravity; Calculate the statistical indicator characteristics of each axis obtained; Calculate the first fluctuation intensity of each axis as a whole, and then calculate the second fluctuation intensity of each axis as a whole for the data rotated to the standard posture; The first information entropy, the first range, the second information entropy, the second range, the statistical indicator feature, the first fluctuation severity, and the second fluctuation severity are combined, and a normalization operation is performed; The normalized features are used as training data, and the autoencoder structure is used to encode and decode the data. The first scoring model of the corresponding user is obtained by optimizing the loss, and the loss function is the mean square error loss function.

4. The traffic accident detection model optimization method according to claim 2, characterized in that: The method for establishing the second scoring model specifically includes: Obtain the three-axis acceleration time series data blocks of all historical minor accident data as training samples; Cut each three-axis acceleration time series data block into several parts in chronological order; Calculate the third information entropy and the third range for each piece of data, and then calculate the fourth information entropy and the fourth range for the data rotated to a standard posture, where the standard posture is a device posture with x and y axes at 0 and z axis at 1g in a static state, where g is the acceleration of gravity; Calculate the statistical indicator characteristics of each axis obtained; Calculate the third fluctuation intensity of each axis as a whole, and then calculate the fourth fluctuation intensity of each axis as a whole for the data rotated to the standard posture; The third information entropy, the third range, the fourth information entropy, the fourth range, the statistical indicator feature, the third fluctuation severity and the fourth fluctuation severity are combined and normalized; The normalized features are used as training data, and the autoencoder structure is used to encode and decode the data. The second scoring model of the corresponding user is obtained by optimizing the loss, and the loss function is the mean square error loss function.

5. The traffic accident detection model optimization method according to claim 3 or 4, characterized in that: The statistical indicator characteristics include standard deviation, 10th quantile, 25th quantile, 50th quantile, 75th quantile, 90th quantile and mean.

6. The traffic accident detection model optimization method according to claim 1, characterized in that: The preset weight is obtained by quantifying the analysis capabilities of the historical missed detections of the first score and the second score.

7. The traffic accident detection model optimization method according to claim 6, characterized in that: The quantification of the analysis capability of the historical missed detection of the first score and the second score specifically includes: w1=l1 / (l1+l2), w2=l2 / (l1+l2), where w1 is the weight of the first score, w2 is the weight of the second score, l1 is the number of successful analyses of historical missed accidents using the first score, and l2 is the number of successful analyses of historical missed accidents using the second score.

8. A traffic accident detection model optimization device, characterized in that: The device comprises: A first information acquisition module, which acquires user history data within a preset period of time and missed detection data in the user history data that is not correctly identified as an accident; A first anomaly scoring module, wherein the first anomaly scoring module establishes an anomaly scoring model with the user historical data to obtain a first scoring model, and estimates a time series anomaly score of each of the missed data with the first scoring model to obtain a first score; A second information acquisition module, the second information acquisition module historical minor accident data; A second anomaly scoring module, wherein the second anomaly scoring module establishes an anomaly scoring model based on the historical minor accident data to obtain a second scoring model, and estimates a time series anomaly score of each missed detection data based on the second scoring model to obtain a second score; A third anomaly scoring module, wherein the third anomaly scoring module calculates a weighted score of the first score and the second score with a preset weight to obtain a third score; A model optimization module, wherein the model optimization module sorts the third scores, selects the driving data contained in a preset number of missed detection data with scores sorted from high to low as new training samples, trains the accident detection model, and obtains an optimized accident detection model.

9. A computer device, characterized in that: The computer device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the traffic accident detection model optimization method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which is loaded and executed by a processor to implement the traffic accident detection model optimization method as claimed in any one of claims 1 to 7.

Images