Video-based person re-Identi fication

Abstract

Person Re-identi fication problem aims at matching the same person in non-overlapping cameras. Earlier, most of the person re-identi fication problems were focused on image-based solutions. But with the increase in surveillance and cameras, video-based solutions are used. The video-based solution gives a better result for person re identifi cation as it includes spatial and temporal information of the person which is not present in the single image. In this project, I worked on two models. First, two-stream convolutional networks (TSF-CNN) for extracting spatial and temporal features in videos. I evaluated this model on UCF101 Dataset. Second, I proposed a model using Spatial-Temporal Attention network, TSF-CNN and Attribute network for video-based person re-identifi cation. The TSF-CNN network learns the spatial and temporal features whereas the attribute network learns the attribute of the person. I evaluated this model on MARS and iLIDS-VID Dataset.