A kind of target tracking method and device based on similar template update

Abstract

The present invention provides a target tracking method and device based on similar template update. The method includes: performing image feature extraction on a video frame picture at an initial moment through a target tracking module to obtain an initial model and an initial incremental update model; The new model calculates the incremental update model at time t according to the new model at time t-1 and the incremental update model at time t-1 Calculate the similarity δ between the new model and init , calculate the similarity δ between the new model and incre , according to the similarity δ init and similarity δ incre The new model or the final model at time t is selected by the model update strategy. The present invention calculates and T by using the convolution response init The similarity between , and the final model at time t is selected according to the similarity, which can quickly detect the reliability of the new model.

Description

technical field [0001] The present invention relates to the technical field of image processing, and in particular, to a target tracking method and device based on similar template update. Background technique [0002] Artificial intelligence is an important driving force for a new round of scientific and technological revolution and industrial transformation. As one of the important research directions of artificial intelligence technology in computer vision, target tracking is to detect one or several objects from a video. know the exact location of the target. As current computer vision tasks pay more and more attention to video analysis, object tracking algorithms have received more and more attention. [0003] The target tracking system can be roughly divided into three modules: video frame input, target tracking and result display. The video frame input module is used to read the video data and send it to the target tracking module by frame. The target tracking modu...

AI Technical Summary

Problems solved by technology

However, the performance of traditional features is not accurate enough compared to deep learning features. Because the description ability is not strong enough, and the semantic information of the target image cannot be expressed, it is easier to lose the target when using traditional features to determine the target position.

[0023] 2. The model degradation problem of the correlation filter itself

However, as the target tracking proceeds, the correlation filter is more prone to model degradation, and the model degradation will gradually make the model inaccurate, resulting in the loss of target tracking

Images