CN-121982658-A - Vehicle reverse running detection method based on 3D boundary frame

CN121982658ACN 121982658 ACN121982658 ACN 121982658ACN-121982658-A

Abstract

The invention discloses a vehicle retrograde detection method based on a 3D boundary frame, which relates to the technical field of computer vision target detection, and comprises the following steps of predicting a plurality of key point coordinates of a target vehicle through a 3D target detection model according to a road traffic scene image; the method comprises the steps of constructing a 3D boundary frame of a target vehicle based on a plurality of key point coordinates, determining a first plane area where a head of the target vehicle is located and a second plane area where a tail of the target vehicle is located according to the 3D boundary frame, determining a running direction vector of the target vehicle according to a center point coordinate of the first plane area and a center point coordinate of the second plane area, carrying out vector product operation on a preset lane direction vector and the running direction vector, and judging whether the target vehicle is in reverse running according to an operation result. According to the invention, the vehicle reverse running rapid and accurate judgment is realized based on the 3D bounding box detection of the single frame image, so that the detection instantaneity is improved and the multi-scene adaptability is enhanced.

Inventors

Ju Hewei
XU HAODONG
GAO ZHENGHUA

Assignees

南京交控积图网络科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251202

Claims (10)

1. A vehicle reverse travel detection method based on a 3D bounding box, comprising: S1, predicting a plurality of key point coordinates of a target vehicle through a 3D target detection model according to a road traffic scene image, wherein the plurality of key point coordinates represent a three-dimensional structure of the target vehicle; s2, constructing a 3D boundary frame of the target vehicle based on the plurality of key point coordinates, and determining a first plane area where the head of the target vehicle is located and a second plane area where the tail of the target vehicle is located according to the 3D boundary frame; S3, determining a running direction vector of the target vehicle according to the center point coordinates of the first plane area and the center point coordinates of the second plane area; And S4, carrying out vector product operation on a preset lane direction vector and the running direction vector, and judging whether the target vehicle is in reverse running or not according to an operation result.
2. The vehicle reverse running detection method based on a 3D bounding box according to claim 1, wherein determining, according to the 3D bounding box, a first planar area where a head of the target vehicle is located and a second planar area where a tail of the target vehicle is located, specifically includes: Numbering the plurality of key point coordinates according to preset conditions of the head and tail of the target vehicle, wherein the plurality of key point coordinates comprise at least 8 key point coordinates; the coordinates from the key point 1 to the key point 4 represent the headstock rectangular corner of the 3D boundary box so as to determine the first plane area; The keypoint 5 coordinate to the keypoint 8 coordinate characterize the tailstock rectangular corner of the 3D bounding box to determine the second planar region.
3. The vehicle retrograde detection method based on the 3D bounding box according to claim 1, wherein the 3D object detection model is obtained by training a 3D object detection data set based on YOLOv a 7-Pose key point detection algorithm, and the process of obtaining the 3D object detection data set specifically includes: Acquiring image data containing vehicle running images shot by a plurality of heights and a plurality of pitching angles; Marking a 3D boundary box on a corresponding area of the vehicle running image in the image data by using a marking tool; And exporting and converting the marked image data into a YOLO-Pose format to serve as a 3D target detection dataset.
4. The vehicle reverse travel detection method based on the 3D bounding box according to claim 1, wherein determining whether the target vehicle is in reverse travel according to the operation result specifically comprises: The vector product operation specifically comprises the steps of calculating a quantity product of a preset lane direction vector and the running direction vector, judging that the target vehicle is forward running when the result of the quantity product is larger than 0, and judging that the target vehicle is reverse running when the result of the quantity product is smaller than 0.
5. The vehicle retrograde detection system based on the 3D bounding box is characterized by comprising a target detection module, a 3D construction module, a vector calculation module and a retrograde judgment module; the target detection module is used for predicting a plurality of key point coordinates of a target vehicle through a 3D target detection model according to the road traffic scene image, wherein the plurality of key point coordinates represent a three-dimensional structure of the target vehicle; the 3D construction module is used for constructing a 3D boundary frame of the target vehicle based on the plurality of key point coordinates, and determining a first plane area where the head of the target vehicle is located and a second plane area where the tail of the target vehicle is located according to the 3D boundary frame; the vector calculation module is used for determining a running direction vector of the target vehicle according to the center point coordinates of the first plane area and the center point coordinates of the second plane area; the reverse determination module is used for performing vector product operation on a preset lane direction vector and the running direction vector, and determining whether the target vehicle reverses according to an operation result.
6. The vehicle reverse running detection system based on a 3D bounding box according to claim 5, wherein determining, according to the 3D bounding box, a first planar area where a head of the target vehicle is located and a second planar area where a tail of the target vehicle is located, specifically includes: Numbering the plurality of key point coordinates according to preset conditions of the head and tail of the target vehicle, wherein the plurality of key point coordinates comprise at least 8 key point coordinates; the coordinates from the key point 1 to the key point 4 represent the headstock rectangular corner of the 3D boundary box so as to determine the first plane area; The keypoint 5 coordinate to the keypoint 8 coordinate characterize the tailstock rectangular corner of the 3D bounding box to determine the second planar region.
7. The vehicle retrograde detection system based on 3D bounding box according to claim 5, wherein the 3D object detection model is obtained by training a 3D object detection data set based on YOLOv-Pose key point detection algorithm, and the process of obtaining the 3D object detection data set specifically includes: Acquiring image data containing vehicle running images shot by a plurality of heights and a plurality of pitching angles; Marking a 3D boundary box on a corresponding area of the vehicle running image in the image data by using a marking tool; And exporting and converting the marked image data into a YOLO-Pose format to serve as a 3D target detection dataset.
8. The 3D bounding box-based vehicle reverse travel detection system according to claim 5, wherein determining whether the target vehicle is traveling in reverse or not according to the operation result, specifically comprises: The vector product operation specifically comprises the steps of calculating a quantity product of a preset lane direction vector and the running direction vector, judging that the target vehicle is forward running when the result of the quantity product is larger than 0, and judging that the target vehicle is reverse running when the result of the quantity product is smaller than 0.
9. A computer device comprising a processor coupled to a memory, the memory having stored therein at least one computer program that is loaded and executed by the processor to cause the computer device to implement the method of any of claims 1-4.
10. A computer readable storage medium having stored therein at least one computer program that is loaded and executed by a processor to cause a computer to implement the method of any one of claims 1 to 4.

Description

Vehicle reverse running detection method based on 3D boundary frame Technical Field The invention relates to the technical field of computer vision target detection, in particular to a vehicle reverse running detection method based on a 3D boundary box. Background With the development of intelligent traffic systems, vehicle retrograde detection is receiving attention as a key technology for guaranteeing road traffic safety. The prior art mainly faces two major challenges, namely extremely high real-time requirement, requirement of detection and early warning triggering in millisecond level, and complex scene adaptability problem, including road side cameras with different installation heights, unmanned aerial vehicle visual angles with changeable pitching angles and invisible conditions of partial key points caused by shielding of vehicles or other targets. Together, these factors make it difficult for existing methods to meet the requirements of detection accuracy, response speed and scene generalization capability at the same time. For a road side fixed camera scene, a detection scheme based on video streaming is commonly adopted in the prior art. According to the scheme, the position of the vehicle is obtained by utilizing a target detection algorithm, the same vehicle in continuous multi-frame images is associated by combining a target tracking algorithm, a running track is generated, and the reverse running behavior is judged by comparing the preset lane direction with the vehicle track vector. In an unmanned aerial vehicle inspection scene, the existing method calculates the vehicle running direction in real time by detecting 4 key feature points on the surface of a vehicle and utilizing vector operation, so that retrograde judgment is realized under a dynamic view angle. However, the above-described technical route has significant drawbacks. The road side camera scheme depends on a multi-frame tracking mechanism, tracking loss or identity confusion is easy to cause in a dense vehicle or abrupt illumination scene, and the real-time performance is severely restricted by huge calculation overhead introduced by a tracking algorithm. The unmanned plane scheme has strict requirements on the shooting visual angle, and when part of key points of the vehicle are shielded by the self structure or other vehicles, the direction calculation is invalid or a great error is generated. Both methods fail to balance between computational efficiency and robustness, and are difficult to adapt to the requirements of intelligent traffic full scene deployment. In summary, the existing vehicle reverse detection technology is excessively dependent on timing tracking or is easily limited by view angle, and has limitations that are difficult to overcome in terms of real-time performance, accuracy and scene adaptability, so that a novel detection method capable of rapidly judging based on a single frame image, independent of full key point visibility and adapting to multi-view input is needed. Disclosure of Invention The invention aims to solve the technical problems of poor real-time performance caused by multi-frame tracking and inaccurate direction judgment caused by key point shielding in the vehicle reverse detection method aiming at the defects of the prior art, and particularly provides a vehicle reverse detection method based on a 3D bounding box, which comprises the following steps: 1) In a first aspect, the present invention provides a vehicle reverse running detection method based on a 3D bounding box, and the specific technical scheme is as follows: S1, predicting a plurality of key point coordinates of a target vehicle through a 3D target detection model according to a road traffic scene image, wherein the plurality of key point coordinates represent a three-dimensional structure of the target vehicle; s2, constructing a 3D boundary frame of the target vehicle based on the plurality of key point coordinates, and determining a first plane area where the head of the target vehicle is located and a second plane area where the tail of the target vehicle is located according to the 3D boundary frame; S3, determining a running direction vector of the target vehicle according to the center point coordinates of the first plane area and the center point coordinates of the second plane area; And S4, carrying out vector product operation on a preset lane direction vector and the running direction vector, and judging whether the target vehicle is in reverse running or not according to an operation result. The vehicle retrograde detection method based on the 3D bounding box has the following beneficial effects: The three-dimensional structure key points of the vehicle are directly predicted from a single frame image through the 3D boundary frame to determine the driving direction, multi-frame tracking calculation cost is avoided, the real-time performance of retrograde detection is improved, the direction vector is calculated by