Research On Video Target Retrieval Method And Application Based On Region Of Interest

Video target retrieval technology is a technology that uses computer vision algorithms to determine whether a specific target exists in a video.This technology has a wide range of application prospects,such as intelligent security,unmanned supermarkets,and human-computer interaction.track.In recent years,the benefits of deep learning algorithms have progressed,and target retrieval methods have developed rapidly.However,most of the related work is carried out on an image-by-image basis.Few people consider the target retrieval task with video as the starting point;In the end-to-end target retrieval algorithm,there is always a problem that the loss function is difficult to converge due to the mutual interference of the detection loss and the re-identification loss,and the problem also limits the development of the current algorithm.This paper proposes a set of cascaded key frame extraction methods based on practical application scenarios,which should deal with the problem of high video data storage space and low signal-to-noise ratio.At the same time,a video object retrieval algorithm based on region of interest is designed to solve the ills that the loss function is difficult to converge in related research.The specific main research results are as follows:(1)Propose a cascaded key frame extraction method for video target retrieval tasks.This paper investigates the existing key frame extraction technology,and designs a segmentation based on edge contour change rate and a cascaded key frame extraction algorithm based on hierarchical clustering for retrieving video application scenarios.Experiments show that the key frame extraction of this paper is designed.The algorithm removes most of the redundant information and compresses the storage cost of the video content to the previous 23%.(2)Propose a target retrieval algorithm based on region of interest to solve the problem that retrieval loss is difficult to converge.The algorithm designed in this paper draws on the idea of candidate frame extraction in RPN network,and introduces a new weighting mechanism of region of interest for verification model training.Experiments show that the F1 value has increased from 53.24 to 79.35 compared with the most advanced retrieval algorithms.The problem that the retrieval loss is difficult to converge is effectively solved,and the target positioning accuracy and the retrieval accuracy are improved.In addition,this paper also specially designed the feature dimension reduction technology,which can further reduce the characteristic storage cost of the network output,thus supporting online level instance retrieval.Finally,this paper builds a related data set for the practical application scenarios of the nature reserve cross-camera pedestrian search,and applies the model framework proposed in this paper,and successfully implements end-to-end pedestrian search on the algorithm.It can be foreseen that the application of this technology can effectively improve the staff s ability to monitor and manage nature reserves.