Person Re-Identification using Background Subtraction and Siamese Network for Pose Varians

Abstract

Person Re-Identification is a process where the algorithm in charge of matching the similarity of two objects. This method can be used as an alternative solution for the current traditional security surveillance. Many modern technologies that use this model, especially in the use of Video Surveillance. The expected output from the use of this model is the process of monitoring and detecting the similarity of two human objects more efficiently and accurately. However, in its implementation there are still many problems found by previous researchers related to Person Identification. Some of the problems that are often encountered in re-identification are image occlusion, pose variance, illuminati, etc. One of the problems that occur is the difference in poses, the difference in poses causes the re-identification process to often experience errors because the features obtained by the two images may experience differences. In this study, trying to implement the algorithm on a video dataset. There is an additional preprocessing which uses the image segmentation method to extract objects from the video dataset. After pre-processing, the image obtained will be re-identified using the Siamese Network Algorithm. The test results obtained an accuracy of 51% and 54% for each architecture. While the accuracy value of object detection obtained is 0.359 and 0.378, which means that the addition of segmentation using the background subtraction model when compared to previous studies is still not effective in dealing with the problem of different poses.