Search

KR-20260063339-A - SYSTEM AND METHOD FOR ANALYSISING TRAFFIC ACCIDENT VIDEO

KR20260063339AKR 20260063339 AKR20260063339 AKR 20260063339AKR-20260063339-A

Abstract

The present invention discloses a traffic accident video analysis system and method. According to a specific implementation example of the present invention, by utilizing a keyframe extraction model of a pre-established deep neural network structure and a Temporal Segment Network (TSN) model to quantify the fault ratio of a traffic accident based on the type of accident scene, characteristics of the accident scene, and the direction of behavior of the accident victim, the traffic accident ratio can be objectively determined, thereby reducing disputes caused by unnecessary traffic accident ratios.

Inventors

  • 문일영
  • 이서영
  • 유연휘
  • 박효경
  • 박병주

Assignees

  • 한국기술교육대학교 산학협력단

Dates

Publication Date
20260507
Application Date
20241030

Claims (7)

  1. An accident scene image extraction unit that extracts keyframes from multiple input images based on a pre-established keyframe extraction model and then outputs an accident scene type from the images in keyframe units; A customer behavior recognition unit that receives the input multiple images and the accident scene type, learns based on a pre-established TSN (Temporal Segment Network) model to derive the accident scene features and the locations of customer A and B, and outputs the behavior direction of customer A and B based on the locations of customer A and B; and A traffic accident video analysis system characterized by including an accident rate determination unit that extracts an accident rate from a pre-established AI model using the above-mentioned accident scene type, accident scene characteristics, and behavioral direction of customer entities A and B as input.
  2. In paragraph 1, the above customer behavior recognition unit is, A preprocessing module that is performed in the pipeline of a TSN model, preprocesses input segment-unit images to a predetermined resolution, standardizes them, and then augments them; A recognizer that extracts a feature map input into a pre-established backbone network (ResNet-50) from the above-mentioned preprocessed segment-unit image; A localization module that extracts snippets from the above-mentioned preprocessed consecutive frame-unit images, combines the extracted snippets to output action recognition, and then outputs the start and end frames of the corresponding action; and A traffic accident video analysis system comprising a classifier that matches the dataset of CLS(Classification)_Head, defined based on the above-mentioned extracted feature map and the behavioral direction of the customer body, and stores it in the form of labels.
  3. In paragraph 2, the localization module is, A traffic accident video analysis system that further includes the time interval of the corresponding action.
  4. In a traffic accident video analysis method performed based on the traffic accident video analysis system of paragraph 1, At least one processor included in the above-mentioned traffic accident video analysis system is, An accident scene image extraction step that extracts keyframes from multiple input images based on a pre-established keyframe extraction model and then outputs an accident scene type from the images in keyframe units; A customer behavior recognition step that receives the input multiple images and the accident scene type, learns based on a pre-established TSN (Temporal Segment Network) model to derive the accident scene features and the locations of customer entities A and B, and outputs the behavioral direction of customer entities A and B based on the locations of customer entities A and B; and A method for analyzing traffic accident video, characterized by including an accident rate determination step that extracts an accident rate from a pre-established AI model using the above-mentioned accident scene type, accident scene characteristics, and behavioral direction of customer entities A and B as input.
  5. In paragraph 4, the above customer behavior recognition step is, A step performed in the pipeline of the above TSN model, wherein the input segment-unit image is preprocessed to a predetermined resolution, standardized, and then augmented; A step of extracting a feature map input into a pre-constructed backbone network (ResNet-50) from the above-mentioned preprocessed segment-unit images; A step of extracting snippets from the above-mentioned preprocessed consecutive frame-unit images, combining the extracted snippets to output action recognition, and then outputting the start and end frames of the corresponding action; and A method for analyzing traffic accident images characterized by including the step of matching the extracted feature map and the behavioral direction of the customer body with a dataset of CLS (Classification)_Head defined based on the above, and saving it in the form of labels.
  6. A computer-readable recording medium characterized by having a program recorded thereon for executing the traffic accident video analysis method of paragraph 4 or 5 on a computer.
  7. In a computer program stored on a computer-readable recording medium for executing a traffic accident video analysis method on a computer in combination with a computer, The above traffic accident video analysis method is, An accident scene image extraction step that extracts keyframes from multiple input images based on a pre-established keyframe extraction model and then outputs an accident scene type from the images in keyframe units; A customer behavior recognition step that receives the input multiple images and the accident scene type, learns based on a pre-established TSN (Temporal Segment Network) model to derive the accident scene features and the locations of customer entities A and B, and outputs the behavioral direction of customer entities A and B based on the locations of customer entities A and B; and An operation program for a traffic accident video analysis system characterized by including an accident rate determination step that extracts an accident rate from a pre-established AI model using the above-mentioned accident scene type, accident scene characteristics, and behavioral direction of customer entities A and B as input.

Description

System and Method for Analyzing Traffic Accident Video The present invention relates to a traffic accident video analysis system and method, and more specifically, to a technology that can quantify and provide the fault ratio of a traffic accident based on the type of accident scene, characteristics of the accident scene, and the behavioral direction of the accident victim using a keyframe extraction model of a pre-established deep neural network structure and a Temporal Segment Network (TSN) model. Research on autonomous vehicles is actively underway. With the emergence of autonomous vehicles, a transitional period in which conventional vehicles and autonomous vehicles coexist is expected to come, and accident rates are expected to increase during this transitional period. Currently, when a traffic accident occurs, the fault ratio is measured according to the General Insurance Association of Korea’s ‘Standards for Recognizing Fault Ratios in Automobile Accidents.’ However, as the cost and reliability of investigating the type of accident are very high, there is a growing trend of disputes over fault ratios requesting reconsideration even in cases where the fault ratio has already been determined. Furthermore, since these disputes often lead to lawsuits and court rulings, the additional resources consumed are substantial. The following drawings attached to this specification illustrate preferred embodiments of the present invention and serve to further enhance understanding of the technical concept of the present invention together with the detailed description of the invention provided below; therefore, the present invention should not be interpreted as being limited only to the matters described in such drawings. Figure 1 is a configuration diagram of a traffic accident video analysis system of one embodiment. Figure 2 is a detailed configuration diagram of the customer behavior recognition unit of Figure 1. Figure 3 is an example diagram showing the output image of the preprocessing module of Figure 2. Figure 4 is an example diagram showing the CLS_Head of the recognizer of Figure 2. Figure 5 is an example diagram showing the accident ratio of the accident ratio judgment unit of Figure 1. Figure 6 is an overall flowchart showing the traffic accident video analysis process of another embodiment. Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals. One embodiment described below specifically explains a traffic accident video analysis system and method. FIG. 1 is a configuration diagram of a traffic accident video analysis system according to one embodiment, FIG. 2 is a detailed configuration diagram of the customer behavior recognition unit of FIG. 1, FIG. 3 is an example diagram showing the output image of the preprocessing module of FIG. 2, FIG. 4 is an example diagram showing the CLS_Head of the recognizer of FIG. 2, and FIG. 5 is an example diagram showing the accident ratio of the accident ratio determination unit of FIG. 1. Referring to FIG. 1 to FIG. 5, the traffic accident video analysis system according to one embodiment is configured to quantify and provide the fault ratio of a traffic accident based on the accident scene type, accident scene characteristics, and the behavior direction of the customer using a keyframe extraction model of a pre-established deep neural network structure and a pre-established TSN (Temporal Segment Network) model. Referring to FIG. 1, the system of one embodiment may include an accident scene video extraction unit (100), a customer behavior recognition unit (200), and an accident ratio determination unit (300). Here, the accident scene video extraction unit (100) extracts key frames of a plurality of collected videos and then extracts an accident scene video by a key frame extraction model of a deep neural network for the extracted key frames. Here, the accident scene video extraction unit100) extracts keyframes of multiple images input through a camera using an API provided by Katna, a deep neural network-based keyframe extraction library, and then performs transfer learning on a pre-trained keyframe extraction model using the extracted keyframe images as input to output an accident scene type. For example, the deep neural network image classification model is a ResNet-18 composed of 18 layers, which is pre-trained on ImageNet 1K. Here, the accident scene type can be set to four types, including parking lots, roadways/places excluding roadways, roundabouts, and T-junctions. Meanwhile, the customer behavior recognitio