Training video generation method for reconstructing traffic accident scene

Abstract

The invention discloses a training video generation method for reconstructing a traffic accident scene, belongs to the technical field of intelligent traffic, and can solve the problem that the pertinence and effectiveness of driving training are reduced because the conventional scene reconstruction method is difficult to reflect typical characteristics of a real traffic accident scene. The method comprises the following steps: S1, generating a plurality of accident samples according to a plurality of traffic accident video data, and extracting a plurality of scene features of each accident sample; s2, performing scene classification on all accident samples according to scene features to obtain a plurality of scene categories and corresponding scene sample sets; s3, determining difference significant features from the scene features according to the scene sample set, and determining key influence features of each scene category from sub-features of the difference significant features; and S4, forming a scene reconstruction data set by using all the key influence characteristics of each scene category, and generating a training video of the corresponding scene category according to the scene reconstruction data set. The method is used for reconstructing the traffic accident scene.

Description

Technical Field

[0001] This invention relates to a method for generating training videos for reconstructing traffic accident scenarios, belonging to the field of intelligent transportation technology. Background Technology

[0002] Vulnerable road users (VRUs) are defined as road users who lack the protection of a metal shell and are vulnerable to collision injuries. They typically include pedestrians, cyclists, e-bike riders, and motorcyclists.

[0003] Currently, researchers have proposed various technical solutions to improve drivers' ability to identify dangerous scenarios and their situational awareness. Some studies summarize the characteristics of typical accident scenarios through accident database analysis and statistics, but this method relies too heavily on existing database resources and struggles to reflect new traffic environment characteristics in a timely manner. Other studies utilize expert experience to construct driver training scenarios, but this method is highly subjective, lacks data support, and fails to accurately reflect the dangerous characteristics of real traffic environments. Still other studies attempt to generate training scenarios based on traffic simulation software. While this method can generate diverse scenarios, it often lacks in-depth analysis of the characteristics of real accident scenarios, resulting in insufficient realism and typicality of the generated scenarios.

[0004] Despite the progress made in driver training, several issues remain to be addressed. First, the lack of systematic analysis of VRU accident scenarios within my country's specific road traffic environment makes it difficult to tailor existing training content to localized traffic characteristics. Second, existing scenario construction methods fail to reflect the typical characteristics of real-world accident scenarios, reducing the relevance and effectiveness of training. Third, scenario reconstruction methods lack theoretical support, making it difficult to guarantee the effectiveness of reconstructed scenarios. Finally, existing driver training systems struggle to provide targeted training for hazardous scenarios, hindering the improvement of drivers' hazard avoidance abilities. Summary of the Invention

[0005] This invention provides a training video generation method for reconstructing traffic accident scenarios, which can solve the problem that existing scenario reconstruction methods are unable to reflect the typical characteristics of real traffic accident scenarios, thereby reducing the pertinence and effectiveness of driver training.

[0006] This invention provides a method for generating training videos for reconstructing traffic accident scenes, the method comprising: S1. Generate multiple accident samples based on multiple traffic accident video data of the target area, and extract multiple scene features for each accident sample; each scene feature includes multiple sub-features; S2. Based on the scene characteristics, hierarchical clustering method is used to classify all accident samples into scenes, resulting in multiple scene categories and scene sample sets corresponding to each scene category; S3. Based on the scene sample set, identify significantly different features from multiple scene features, and identify key influencing features for each scene category from the sub-features of each significantly different feature; S4. Combine all key influencing features of each scene category into a scene reconstruction dataset for that scene category, and generate training videos for that scene category based on the scene reconstruction dataset.

[0007] Optionally, S1 specifically includes: Multiple traffic accident video data points were acquired in the target area, and each traffic accident video data point was preprocessed to obtain multiple accident samples; Each accident sample is labeled with a scene to extract multiple scene features for each accident sample.

[0008] Optionally, S2 specifically includes: Each scene feature of each accident sample is transformed into a dummy variable to obtain the classification matrix of each accident sample; Based on the classification matrix of all accident samples, hierarchical clustering is used to classify all accident samples into scenarios, resulting in multiple scenario categories and a scenario sample set corresponding to each scenario category.

[0009] Optionally, in step S3, determining significantly different features from multiple scene features based on the scene sample set specifically includes: Based on the number of samples corresponding to the sub-features of each scene feature in the scene sample set, construct the observation matrix for each scene feature; A chi-square test analysis was performed on the observation matrix of each scene feature to obtain the confidence level of each scene feature, and the scene features corresponding to the confidence levels that meet the preset conditions were taken as the significantly different features.

[0010] Optionally, based on the number of samples corresponding to the sub-features of each scene feature in the scene sample set, an observation matrix for each scene feature is constructed, specifically including: Using the number of rows of each scene feature's sub-features as the number of rows and the number of columns of each scene category as the number of samples in the corresponding scene category of each scene feature in the scene sample set, an observation matrix for each scene feature is constructed.

[0011] Optionally, the key influencing features for each scene category determined from the sub-features of each significantly different feature in S3 specifically include: Obtain the sample percentage of each sub-feature of each significantly different feature in each scene category; When a sub-feature accounts for more than 50% of the samples, the sub-feature is used as the key influencing feature of the corresponding scene category; When the sample proportion of all sub-features is less than or equal to 50%, the two sub-features with the largest sample proportion are taken as the key influencing features of the corresponding scene category.

[0012] Optionally, when a sub-feature accounts for more than 50% of the samples, the sub-feature is used as the key influencing feature of the corresponding scene category, specifically including: When there is a sub-feature with a sample share greater than 50%, and the sample share of the remaining sub-features is less than or equal to 30%, the sub-feature with a sample share greater than 50% is taken as the key influencing feature of the corresponding scene category. When there are sub-features with a sample share greater than 50% and sub-features with a sample share greater than 30%, the sub-features with a sample share greater than 30% are all regarded as key influencing features of the corresponding scenario category.

[0013] Optionally, the preprocessing includes any one or more of image synchronization, image denoising, scale normalization, and image enhancement.

[0014] Optionally, the scene features include road geometry features, road grade features, isolation facility features, and environmental features; The sub-features of the road geometry include one-way roads, two-way two-lane roads, two-way four-lane roads, two-way six-lane roads, and two-way six-lane or more roads; The sub-features of the road classification characteristics include urban arterial roads, urban secondary roads, trunk highways, township and village roads, and internal roads; The sub-features of the isolation facilities include hard median barriers, median marking barriers, no median barriers, hard median barriers, median marking-pedestrian barriers, and no median barriers. The sub-features of the environmental characteristics include sunny, cloudy, rainy, snowy, foggy, hazy, daytime, dusk, nighttime with lighting, and nighttime without lighting.

[0015] Optionally, sub-features of each scene feature are classified and coded using numbers. The beneficial effects that this invention can produce include: This invention provides a method for generating training videos to reconstruct traffic accident scenarios. The method extracts features from accident videos, identifies typical accident scenarios using hierarchical clustering, analyzes differences in scenario features to determine key influencing features, and finally generates training videos from a scenario reconstruction dataset composed of these key influencing features for driver training. This invention addresses the problem that existing scenario reconstruction models struggle to reflect the typical characteristics of real traffic accident scenarios, thus reducing the relevance and effectiveness of driver training.

[0016] The present invention provides a training video generation method for reconstructing traffic accident scenarios. Through systematic analysis of traffic accident videos, it extracts multi-dimensional information, including road geometric features, road grade features, isolation facility features, and environmental features. This analysis method based on real data not only ensures the authenticity and reliability of the analysis results but also reflects the characteristics of typical dangerous scenarios in specific road traffic environments in my country, providing a more targeted basis for driver training.

[0017] This invention provides a training video generation method for reconstructing traffic accident scenarios. It employs dummy variable transformation and hierarchical clustering to classify accident scenarios, achieving quantitative analysis of these scenarios. The optimal number of scenario categories is determined through total sum of squares analysis, avoiding the shortcomings of subjective classification in traditional methods. Simultaneously, based on chi-square tests and correspondence analysis, feature variables that significantly influence the classification results are scientifically identified, providing reliable parameter basis for scene reconstruction. Attached Figure Description

[0018] Figure 1 A flowchart of a training video generation method for reconstructing traffic accident scenarios provided in an embodiment of the present invention; Figure 2 This is a classification diagram of gravel in a typical accident scene at an intersection provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the classification results of typical accident scenarios at intersections provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of typical accident scenarios at intersections provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the embodiments, but the present invention is not limited to these embodiments.

[0020] This invention provides a method for generating training videos to reconstruct traffic accident scenarios, such as... Figures 1 to 4 As shown, the method includes: S1. Generate multiple accident samples based on multiple traffic accident video data of the target area, and extract multiple scene features for each accident sample; each scene feature includes multiple sub-features.

[0021] Specifically, it includes: (1) Acquire multiple traffic accident video data in the target area and preprocess each traffic accident video data to obtain multiple accident samples.

[0022] The target area mentioned above refers to the intersection area or road segment area. In practical applications, after acquiring multiple traffic accident video data of a certain road at a certain time period, the multiple traffic accident video data are first divided into two categories according to whether they occurred at an intersection or on a road segment; then, the multiple traffic accident video data of the intersection area and the road segment area are analyzed and processed separately.

[0023] Preprocessing includes any one or more of the following: image synchronization, image denoising, scale normalization, and image enhancement.

[0024] The aforementioned image synchronization ensures the consistency of time stamps for different video data, facilitating subsequent analysis.

[0025] The above image denoising method uses a Gaussian filter to eliminate image noise and improve image quality.

[0026] The aforementioned standardization involves adjusting videos of different resolutions to a standard size.

[0027] The image enhancement process described above involves performing random rotation on the image within the range of -30° to 30°, random scaling within the range of 0.8 to 1.2 times the original size, random adjustment of the image's HSV space by ±20%, and random adjustment of the image contrast by ±30%.

[0028] These preprocessing steps improve the quality and consistency of video data.

[0029] (2) Each accident sample is labeled with a scene to extract multiple scene features of each accident sample.

[0030] The aforementioned scene features include road geometry, road grade, isolation facilities, and environmental features.

[0031] The sub-features of road geometry include one-way roads, two-way two-lane roads, two-way four-lane roads, two-way six-lane roads, and two-way six-lane or more roads; the sub-features of road grade include urban arterial roads, urban branch roads, trunk highways, township and village roads, and internal roads; the sub-features of isolation facilities include hard median barriers, median marking barriers, no median barriers, hard median barriers on the roadside, median marking-pedestrian barriers, and no median barriers on the roadside; the sub-features of environmental characteristics include sunny, cloudy, rainy, snowy, foggy, hazy, daytime, dusk, nighttime with lighting, and nighttime without lighting.

[0032] Furthermore, the sub-features of each scene feature are classified and coded using numbers.

[0033] In this embodiment of the invention, the specific scene features extracted include: Road geometric features: Classified and coded according to the number of lanes and direction of the road, including one-way road as 1, two-way two-lane road as 2, two-way four-lane road as 3, two-way six-lane road as 4, and two-way six-lane or more road as 5; Road classification characteristics: Classified according to road function and scale, with urban arterial roads (total number of lanes not less than six) as 1, urban branch roads (total number of lanes less than six) as 2, trunk highways as 3, township and village roads as 4, and internal roads (roads within parking lots, residential areas, etc.) as 5. Characteristics of isolation facilities: They are divided into two categories: median barriers and roadside barriers. Median barriers are classified as follows: 1 for hard median barriers, 2 for marked median barriers, and 3 for no median barriers. Roadside barriers are classified as follows: 1 for hard median barriers, 2 for marked median barriers and pedestrian barriers, and 3 for no median barriers. Environmental characteristics: including weather conditions and lighting conditions. Weather conditions are categorized as follows: sunny (1), cloudy (2), rain, snow, fog, or haze (3); lighting conditions are categorized as follows: daytime (1), dusk (2), nighttime with lighting (3), and nighttime without lighting (4).

[0034] S2. Based on the scene characteristics, hierarchical clustering is used to classify all accident samples into scenes, resulting in multiple scene categories and scene sample sets corresponding to each scene category.

[0035] Specifically, this includes: first, transforming each scene feature of each accident sample into a dummy variable to obtain a classification matrix for each accident sample; then, based on the classification matrix of all accident samples, using hierarchical clustering to classify all accident samples into scenes, resulting in multiple scene categories and a scene sample set corresponding to each scene category.

[0036] Among them, dummy variable transformation is applied to each scene feature, when the sample Having the first When individual features are present, ,otherwise This yields the classification matrix for each accident sample.

[0037] Then, hierarchical clustering is used to determine the number of scene categories and the scene sample set corresponding to each scene category.

[0038] S3. Based on the scene sample set, identify the significantly different features from multiple scene features, and identify the key influencing features of each scene category from the sub-features of each significantly different feature.

[0039] Based on the scene sample set, the above-mentioned significantly different features were identified from multiple scene features, specifically including: (1) Construct the observation matrix of each scene feature based on the number of samples corresponding to the sub-features of each scene feature in the scene sample set.

[0040] Specifically, the observation matrix for each scene feature is constructed by using the number of rows of each scene feature's sub-features and the number of columns of each scene category, based on the number of samples of each scene feature's sub-features in the corresponding scene category in the scene sample set.

[0041] (2) Perform chi-square test analysis on the observation matrix of each scene feature to obtain the confidence level of each scene feature, and take the scene feature corresponding to the confidence level that meets the preset conditions as the significant difference feature.

[0042] The above-mentioned preset conditions are pre-set confidence thresholds. The specific value of the confidence threshold is not limited in the embodiments of the present invention, and those skilled in the art can set it according to the actual situation.

[0043] The key influencing features for each scene category are identified from the sub-features of each significantly different feature, specifically including: Obtain the sample percentage of each sub-feature of each significantly different feature in each scene category; When there is a sub-feature with a sample share greater than 50%, the sub-feature is regarded as the key influencing feature of the corresponding scene category; When the sample proportion of all sub-features is less than or equal to 50%, the two sub-features with the largest sample proportion are taken as the key influencing features of the corresponding scene category.

[0044] Furthermore, when a sub-feature accounts for more than 50% of the samples, it is used as the key influencing feature of the corresponding scene category, specifically including: When there is a sub-feature with a sample share greater than 50%, and the sample share of the remaining sub-features is less than or equal to 30%, the sub-feature with a sample share greater than 50% is taken as the key influencing feature of the corresponding scene category. When there are sub-features with a sample share greater than 50% and sub-features with a sample share greater than 30%, the sub-features with a sample share greater than 30% are all regarded as key influencing features of the corresponding scenario category.

[0045] Based on the above rules, the key impact characteristics of each scenario category are determined to provide a basis for scenario reconstruction.

[0046] S4. Combine all key influencing features of each scene category into a scene reconstruction dataset for that scene category, and generate training videos for that scene category based on the scene reconstruction dataset.

[0047] In practical applications, a local knowledge base can be built using the reconstructed scene dataset, and high-fidelity training videos can be generated in batches based on professional video-generating AI tools such as Sora to meet the specific needs of driver training in different scenarios.

[0048] Furthermore, the present invention also includes a scene dynamic adjustment mechanism, which can adaptively adjust the setting of scene parameters, including road environment complexity, traffic participant behavior patterns and risk event triggering probability, based on the driver's response to the current scene.

[0049] The accident scene reconstruction method of the present invention is applicable to different road levels such as urban roads, rural roads and highways, and vulnerable road users include pedestrians, cyclists, electric bicycle riders and motorcyclists.

[0050] In one specific embodiment, the present invention analyzes accident scenarios at intersections and road segments respectively. By statistically analyzing 178 collision accident videos of passenger vehicles and vulnerable road users in an open-source dataset, 108 sets of intersection collision data were obtained.

[0051] The optimal number of scene categories is determined by calculating the sum of squares of the data within each category. Taking a gravel map of an intersection as an example, refer to... Figure 2 As shown, after category 4, the total sum of squares decreases at a rate approaching zero, and the curve gradually flattens out, indicating that the intersection sample data has the highest stability when divided into 4 scenario categories.

[0052] Reference Figure 3 The scene classification results shown indicate that in intersection accident scenarios, the first scene category (i.e., scene one) accounts for 19%, the second scene category (i.e., scene two) accounts for 20%, the third scene category (i.e., scene three) accounts for 42%, and the fourth scene category (i.e., scene four) accounts for 19%.

[0053] Reference Figure 4 The schematic diagram and Table 1 show the categories and key influencing characteristics of intersection accident scenarios. Different categories of accident scenarios have significant differences in characteristics. Taking intersection scenarios as an example: The first type of scenario mainly occurs on urban side roads, mostly at intersections without traffic lights; The second type of scenario mainly occurs on urban arterial roads, specifically at traffic light intersections; The third category of scenarios includes two types: urban trunk roads and rural roads. The fourth type of scenario mainly occurs on urban main roads.

[0054] Table 1. Intersection Scene Categories and Key Influencing Characteristics

[0055] Compared with the prior art, the beneficial effects of the present invention include: 1. This invention provides a scenario analysis method based on real accident data. Through systematic analysis of traffic accident videos, it extracts multi-dimensional information including road geometry features, road grade, median barriers, roadside barriers, weather conditions, lighting conditions, accident object type and status, and the movement and evasive maneuvers of the accident vehicles. This analysis method based on real data not only ensures the authenticity and reliability of the analysis results but also reflects the typical dangerous scenario characteristics under specific road traffic conditions in my country, providing a more targeted basis for driver training.

[0056] 2. A scientific scene classification system was established. This invention employs hierarchical clustering to classify accident scenes. Through steps such as dummy variable transformation, Euclidean distance calculation, and average linkage analysis, quantitative analysis of accident scenes was achieved. The optimal number of classification categories was determined through total sum of squares analysis, avoiding the shortcomings of subjective division in traditional methods. Simultaneously, based on chi-square tests and correspondence analysis, characteristic variables that significantly influence the classification results were scientifically identified, providing reliable parameter basis for scene reconstruction.

[0057] 3. A dynamic scenario adjustment mechanism has been established. This invention dynamically adjusts the road environment parameters, traffic participant behavior characteristics, and accident risk levels of the reconstructed scenario. It can adaptively adjust scenario parameters, including road environment complexity, traffic participant behavior patterns, and the probability of risk event triggering, based on the driver's training performance. This dynamic adjustment mechanism enables personalized instruction and effectively improves training outcomes.

[0058] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for generating training videos for reconstructing traffic accident scenarios, characterized in that, The method includes: S1. Generate multiple accident samples based on multiple traffic accident video data of the target area, and extract multiple scene features for each accident sample; each scene feature includes multiple sub-features; S2. Based on the scene characteristics, hierarchical clustering method is used to classify all accident samples into scenes, resulting in multiple scene categories and scene sample sets corresponding to each scene category; S3. Based on the scene sample set, identify significantly different features from multiple scene features, and identify key influencing features for each scene category from the sub-features of each significantly different feature; S4. Combine all key influencing features of each scene category into a scene reconstruction dataset for that scene category, and generate training videos for that scene category based on the scene reconstruction dataset.

2. The method according to claim 1, characterized in that, S1 specifically includes: Multiple traffic accident video data points were acquired in the target area, and each traffic accident video data point was preprocessed to obtain multiple accident samples; Each accident sample is labeled with a scene to extract multiple scene features for each accident sample.

3. The method according to claim 1, characterized in that, S2 specifically includes: Each scene feature of each accident sample is transformed into a dummy variable to obtain the classification matrix of each accident sample; Based on the classification matrix of all accident samples, hierarchical clustering is used to classify all accident samples into scenarios, resulting in multiple scenario categories and a scenario sample set corresponding to each scenario category.

4. The method according to claim 1, characterized in that, The step S3, which involves determining significantly different features from multiple scene features based on the scene sample set, specifically includes: Based on the number of samples corresponding to the sub-features of each scene feature in the scene sample set, construct the observation matrix for each scene feature; A chi-square test analysis was performed on the observation matrix of each scene feature to obtain the confidence level of each scene feature, and the scene features corresponding to the confidence levels that meet the preset conditions were taken as the significantly different features.

5. The method according to claim 4, characterized in that, Based on the number of samples corresponding to the sub-features of each scene feature in the scene sample set, an observation matrix for each scene feature is constructed, specifically including: Using the number of rows of each scene feature's sub-features as the number of rows and the number of columns of each scene category as the number of samples in the corresponding scene category of each scene feature in the scene sample set, an observation matrix for each scene feature is constructed.

6. The method according to claim 1, characterized in that, The key influencing features for each scene category determined from the sub-features of each significantly different feature in S3 specifically include: Obtain the sample percentage of each sub-feature of each significantly different feature in each scene category; When a sub-feature accounts for more than 50% of the samples, the sub-feature is used as the key influencing feature of the corresponding scene category; When the sample proportion of all sub-features is less than or equal to 50%, the two sub-features with the largest sample proportion are taken as the key influencing features of the corresponding scene category.

7. The method according to claim 6, characterized in that, When a sub-feature accounts for more than 50% of the samples, the sub-feature is used as the key influencing feature of the corresponding scene category, specifically including: When there is a sub-feature with a sample share greater than 50%, and the sample share of the remaining sub-features is less than or equal to 30%, the sub-feature with a sample share greater than 50% is taken as the key influencing feature of the corresponding scene category. When there are sub-features with a sample share greater than 50% and sub-features with a sample share greater than 30%, the sub-features with a sample share greater than 30% are all regarded as key influencing features of the corresponding scenario category.

8. The method according to claim 2, characterized in that, The preprocessing includes any one or more of the following: image synchronization, image denoising, scale normalization, and image enhancement.

9. The method according to claim 1 or 2, characterized in that, The scene features include road geometry features, road grade features, isolation facility features, and environmental features; The sub-features of the road geometry include one-way roads, two-way two-lane roads, two-way four-lane roads, two-way six-lane roads, and two-way six-lane or more roads; The sub-features of the road classification characteristics include urban arterial roads, urban secondary roads, trunk highways, township and village roads, and internal roads; The sub-features of the isolation facilities include hard median barriers, median marking barriers, no median barriers, hard median barriers, median marking-pedestrian barriers, and no median barriers. The sub-features of the environmental characteristics include sunny, cloudy, rainy, snowy, foggy, hazy, daytime, dusk, nighttime with lighting, and nighttime without lighting.

10. The method according to claim 9, characterized in that, Each scene feature's sub-features are classified and coded using numbers.

Images