Person Re-identification Using Deep Learning with Mask-RCNN

Abstract

Person re-identification comprises of matching pictures of a particular person captured in a matrix of cameras with non-overlapping fields of view. The task is not the same as the classic id and detection tasks. The identification comprises of ascertaining the identity of a person in an image, whereas the detection comprises of rotoscoping people from their background, in the absence of the identity of their knowledge. Person re-identification answers the question if the given image is of the same person as the query image. The task of detection tells us whether it is a person and the task of identification helps to know who the person is. But the task of re-identification tells us when and where this particular person appeared, respective to a given camera, and using multiple cameras, person re-identification permits the assessment of his/her path for a tiny time period. The task of person re-identification task depends on the pedestrian detection task. In step 1, we form a set of pedestrians by collecting the cropped pedestrian images or extract the pedestrian image signature from every camera seen in the matrix. Next, we compare the similarity or the distance from the query image. At the end, we show the closest images based on the measured similarity.It is generally considered that the person keeps the same clothes in multiple scenes. Thus, this method works most effectively in a short time period, which makes sure that the visual appearance of the person remains the same. A huge constraint of re-identification is that we cannot apply it to find similarities between people after some days as their appearance may have changed. This is also partly because biometrics like face identification are not exactly possible from far placed