A real-time video multi-target deduplication method, terminal equipment and storage medium

Abstract

The present invention relates to a real-time video multi-target deduplication method, terminal equipment and storage medium. In the method, the following steps are included: S100: extract pictures of each target in the screen according to the video frame picture at the current moment; S200: according to Get the corresponding structured information data from the pictures of each target: target image feature vector, tracking information and image quality score; S300: Set the set of structured information data of each target before the current moment as the target set, and according to the structure of the target The structured information data of each target is clustered, and the structured information data of each target at the current moment is added to the target set, so that the same target at different times in the target set only appears once, and the target that appears is the target with the best relative quality. The present invention realizes the output result of target detection through clustering, and only one picture of the same object is finally output, which greatly reduces the redundant work of the system, reduces the misjudgment rate of detection to a certain extent, and improves the robustness.

Description

technical field [0001] The present invention relates to the technical field of computer vision, in particular to a multi-object deduplication method for real-time video, a terminal device and a storage medium. Background technique [0002] Object detection is one of the basic tasks in the field of computer vision. Its main job is to identify multiple objects in a picture and locate their positions in the image. The current popular methods for extracting structured information from video mainly rely on the target detection model to mark the target of interest first, and then further analyze the features of each marked target to determine the color, shape and other more specific information of the target. . [0003] The output of the video frame after target detection is directly used as the input of subsequent applications (such as face recognition, license plate recognition), and multiple repeated pictures of the same target will be collected, which is generally unavoidable...

AI Technical Summary

Problems solved by technology

The eigenvectors extracted by traditional feature descriptors (such as: LBP, SIFT, ORB, etc.) have limited accuracy in calculating the similarity of target pictures, and the probability of misjudgment is relatively large even within a certain period of time, but its The advantage is that the amount of calculation for extracting some feature vectors is relatively small and fast; the correlation filtering algorithm calculates the image within a certain range in the next frame through the target detector trained by the previous frame of the target, and selects the position of the image with the largest response to determine it as The object to be tracked, the algorithm involves many complex mathematical operations such as matrix diagonalization, inversion, etc. Many subsequent optimization algorithms use different methods to reduce the amount of matrix calculations to improve the operation speed, but even in an ideal state, it is very difficult It is difficult to meet the real-time tracking of multiple targets that appear at the same time, especially in the case of limited hardware performance; for the algorithm based on the convolutional neural network model, the more complex the network is constructed, the more accurate the distinction between different target pictures is, but Relatively, its computational complexity is getting higher and higher, which is contrary to the original intention of reducing the hardware performance consumption of pictures of the same target at different times, so it is not suitable for current application scenarios

Images