CN-115984741-B - Method and device for detecting scattering objects

CN115984741BCN 115984741 BCN115984741 BCN 115984741BCN-115984741-B

Abstract

The application discloses a method and a device for detecting a cast object, which are used for detecting the cast object through a motion detection model and a deep learning cast object detection model and improving the accuracy of the cast object detection. The application provides a method for detecting a casting object, which comprises the steps of obtaining an acquisition image aiming at a target shooting scene, determining an interested area in the acquisition image, detecting a moving target of the interested area through a moving detection model to obtain at least one moving target image, detecting the interested area and the at least one moving target image through a deep learning casting object target detection model, and determining a casting object detection result corresponding to the acquisition image.

Inventors

Ye Fanjie
CHENG MIAO
ZHOU XIANGMING
ZHANG HEQUN
FU KAI
LV ZHIBING

Assignees

浙江大华技术股份有限公司

Dates

Publication Date: 20260512
Application Date: 20221227

Claims (8)

1. A method of detecting a casting, the method comprising: acquiring an acquisition image obtained aiming at a target shooting scene, and determining an interested region in the acquisition image; detecting a moving target in the region of interest through a motion detection model to obtain at least one moving target image; Performing casting object detection on the region of interest and the at least one moving object image through a deep learning casting object detection model, and determining casting object detection results corresponding to the acquired images; the detecting the moving object of the region of interest by the motion detection model to obtain at least one moving object image includes: Determining the outline of each moving object based on the region of interest, and drawing a motion detection frame on the outline to obtain a motion detection frame of each moving object, wherein an image in the motion detection frame is an image of the moving object; filtering the motion detection frame of each moving object according to a preset motion detection frame of a non-scattering object to obtain at least one moving object image; The step of performing the object detection on the region of interest and the at least one moving object image through the deep learning object detection model, and determining the object detection result corresponding to the acquired image includes: Judging whether a cast image exists in the region of interest by carrying out cast detection on the region of interest; if yes, filtering the at least one moving target image according to the cast object image, and obtaining a cast object detection result corresponding to the acquired image by utilizing the rest moving target image and the cast object image; Otherwise, obtaining a casting object detection result corresponding to the acquired image by using the at least one moving object image.
2. The method of claim 1, wherein the filtering the motion detection frame of each moving object according to the preset motion detection frame of the non-cast object to obtain at least one moving object image comprises: For each motion detection frame, carrying out cross-over ratio with a preset motion detection frame of a non-cast object, and filtering out the motion detection frame when the cross-over ratio is larger than a preset threshold value; And expanding each residual motion detection frame based on the width and height of the original acquired image, and extracting the image of the moving target from the expanded motion detection frames to obtain the moving target image.
3. The method of claim 1, further comprising the step of pre-training the deep-learning casting target detection model as follows: Calculating total loss function values of a preset number of sample images, wherein the total loss function values are the sum of preset category loss function values and comparison coding priori loss function values; And adjusting the weights of the neural networks of each layer in the deep learning casting object target detection model according to the total loss function value.
4. The method according to claim 1, wherein the method further comprises: And judging whether to send out the casting early warning information based on casting detection results corresponding to the acquired images obtained by continuously shooting the scene for the target by a plurality of frames.
5. The method according to claim 4, wherein the determining whether to send out the casting early warning information based on the casting detection result corresponding to the acquired image obtained for the target shooting scene by using the continuous multi-frames includes: when the same cast image exists in the cast object detection results corresponding to the acquired images obtained by continuously shooting the scene for the target for multiple frames, judging whether the images exist in a preset database for storing non-cast object images or not; If so, sending out the pre-warning information of the scattering objects when the same scattering object image is not matched with the images in the database; otherwise, the early warning information of the scattering objects is directly sent.
6. A device for detecting a scattering object, comprising: a memory for storing program instructions; A processor for invoking program instructions stored in said memory to perform the method of any of claims 1-5 in accordance with the obtained program.
7. A computer program product for a computer, characterized in that it comprises software code portions for performing the method according to any of claims 1 to 5 when the product is run on the computer.
8. A computer-readable storage medium storing computer-executable instructions for causing a computer to perform the method of any one of claims 1 to 5.

Description

Method and device for detecting scattering objects Technical Field The application relates to the field of intelligent traffic, in particular to a method and a device for detecting a scattering object. Background Traffic road casting detection is an important point and difficulty in the intelligent traffic field, because traffic road casting not only causes the road to become messy, but also endangers the life safety of drivers in severe cases. In recent years, the situation that the personal safety is threatened due to the fact that objects are thrown on traffic roads frequently happens, and management and control of the actions are increasingly emphasized by traffic authorities around the country. The road traffic throwing behavior can be detected by using a gun camera and a ball camera, is usually judged by adopting a target detection mode, but is limited by the variety of throwing objects, the application environment is greatly disturbed, and the recall rate and the accuracy rate of detection are poor. Disclosure of Invention The embodiment of the application provides a method and a device for detecting a cast object, which are used for detecting the cast object through a motion detection model and a deep learning cast object target detection model and improving the accuracy of the cast object detection. The embodiment of the application provides a method for detecting a scattering object, which comprises the following steps: acquiring an acquisition image obtained aiming at a target shooting scene, and determining an interested region in the acquisition image; detecting a moving target in the region of interest through a motion detection model to obtain at least one moving target image; and performing casting object detection on the region of interest and the at least one moving object image through a deep learning casting object detection model, and determining casting object detection results corresponding to the acquired images. The method comprises the steps of obtaining an acquisition image aiming at a target shooting scene, determining an interested area in the acquisition image, detecting a moving target in the interested area through a moving detection model to obtain at least one moving target image, detecting the interested area and the at least one moving target image through a deep learning cast object detection model, and determining a cast object detection result corresponding to the acquisition image, so that the detection of the cast object is realized through the moving detection model and the deep learning cast object detection model, and the accuracy of cast object detection is improved. In some embodiments, the detecting the moving object of the region of interest by using the motion detection model, to obtain at least one moving object image, includes: Determining the outline of each moving object based on the region of interest, and drawing a motion detection frame on the outline to obtain a motion detection frame of each moving object, wherein an image in the motion detection frame is an image of the moving object; and filtering the motion detection frame of each moving object according to a preset motion detection frame of the non-scattering object to obtain at least one moving object image. By the method, the motion detection frame of the moving object is drawn through the motion detection model, and the motion detection frame of the non-cast object is filtered, so that the detection accuracy is improved. In some embodiments, the filtering the motion detection frame of each moving object according to the preset motion detection frame of the non-scattering object to obtain at least one moving object image includes: For each motion detection frame, carrying out cross-over ratio with a preset motion detection frame of a non-cast object, and filtering out the motion detection frame when the cross-over ratio is larger than a preset threshold value; And expanding each residual motion detection frame based on the width and height of the original acquired image, and extracting the image of the moving target from the expanded motion detection frames to obtain the moving target image. By the method, the motion detection frames are filtered, and images in the rest motion detection frames are extracted, so that the accuracy of subsequent detection is improved. In some embodiments, the performing, by using the deep learning object detection model, object detection on the region of interest and the at least one moving object image, and determining an object detection result corresponding to the acquired image includes: Judging whether a cast image exists in the region of interest by carrying out cast detection on the region of interest; if yes, filtering the at least one moving target image according to the cast object image, and obtaining a cast object detection result corresponding to the acquired image by utilizing the rest moving target image and the cast object image; Otherwise, obtaini