Research On Video Object Re-identification Using Region Edge Histogram

The main task of video object re-identification is to re-identify the signed moving object in unknown video data.This paper firstly reviews the algorithm framework of video object re-identification,and explains the key point of video object re-identification from scene variation and object's number.After that,image feature detection which is the core issue of video object re-identification is analyzed.Then,a video object re-identification algorithm based on region edge histogram is proposed.We start from foreground detection of Vibe.Secondly,foreground images' feature vector of region edge histogram in different color space is extracted.Thirdly,a brightness and color transfer function which is used to make feature transform is proposed.Finally,we do similarity measurement using Euclidean distance and Bhattacharyya distance.Experiments in single scene and multiple scene are made.A video data set with single scene is created which includes color video for forty hours in length.More than 10000 foreground objects are detected.PRID-450 S and 3Dpes are used as multiple scene.The innovation points of the research are as follows:1.The region edge histogram is used as the input feature of the video object re-identification algorithm,and the best feature vector dimension of the region edge histogram is verified by experiments.Then,color region edge histogram in HSV space is proposed to enhance the feature's expression.In gray space,the result of experiment in single scene data set shows that the best feature vector dimension is 32 and the re-identification rate is 72.4%.In HSV space,it also achieves the re-identification rate of 86.7% and improves the rate by about 10% than other histograms.2.A brightness and color transfer function is proposed.After the feature extraction is completed,the feature vector is transformed.The whole brightness and color information of the unknown video object can be changed to match with the object to be detected through the transfer function,so as to eliminate the interference caused by the change of the scene.In multiple scenes of PRID-450 S and 3Dpes,the application of transfer function improves the re-identification rate by about 2%.The final re-identification rates are 44% and 51.5% which are improved by about 2% and 4% than state-of-the-art algorithms.