A Study on Person Re-identification in Multi-view Surveillance Videos

Nowadays, surveillance cameras are widely installed at every edge due to their largescale applications ranging from monitoring to smart city and indoor offices for video conferencing. However, to perform several tasks such as detection, tracking, and identification of persons in video data, it is an essential step to develop an algorithm specifically using a distributed network. For this purpose, numerous techniques have been developed in the field of computer vison for single-view images/video analysis. However, overviewing the literature of video data analytics, the techniques for multi-view analysis are missing on extreme level. Therefore, in this paper, we propose a method for person re-identification and tracking in multi-view images through hybrid approach of several techniques that are usually applied for a single-view image. At first, we generate our own dataset and pass the images into a preprocessing step to detect person using YOLO and subtract the background through Mask-RCNN. Next, the color features are extracted from the images which are fed into K-mean clustering algorithm. We introduce a novel distancing formula by adding weights to Bhattacharyya’s distance to measure the similarity between the clusters of two multi-view input images. The conducted experiments verify and confirm the affectiveness of the proposed method by obtaining 86.23 average precision.