Person re-identification algorithm by image from video surveilance sistem using a neural network compound descriptor

Abstract

For people re-identification in distributed video surveillance systems, it is important to have an algorithm that ensures efficiency when a person is blocked by other people or objects. Therefore, for this task, an algorithm has been developed that involves the compound descriptor formation, which includes a global features vector of a person’s image and three local ones for its upper, middle and down parts. The regions of interest selection is carried out based on the detecting key points results in the image of the human body. If a person's image part is occluded by other people or objects, then it is classified as invisible. The hidden part person image is not used in the formation of the local feature. To obtain it, the average value of the k-nearest neighbors such features of a person’s image is calculated. The experiments performed indicate an increase in re-identification accuracy for the Market-1501, DukeMTMC-ReID, MSMT17 and PolReID1077 datasets.