International Journal of Engineering and Advanced Technology (IJEAT) ISSN: 2249 – 8958, Volume-9 Issue-3, February, 2020 4249 Published By: Blue Eyes Intelligence Engineering & Sciences Publication Retrieval Number: C6524029320 /2020©BEIESP DOI: 10.35940/ijeat.C6524.029320 Abstract: Video Analytics applications like security and surveillance face a critical problem of person re-identification abbreviated as re-ID. The last decade witnessed the emergence of large-scale datasets and deep learning methods to use these huge data volumes. Most current re-ID methods are classified into either image-based or video-based re-ID. Matching persons across multiple camera views have attracted lots of recent research attention. Feature representation and metric learning are major issues for person re-identification. The focus of re-ID work is now shifting towards developing end-to-end re-Id and tracking systems for practical use with dynamic datasets. Most previous works contributed to the significant progress of person re-identification on still images using image retrieval models. This survey considers the more informative and challenging video-based person re-ID problem, pedestrian re-ID in particular. Publicly available datasets and codes are listed as a part of this work. Current trends which include open re-identification systems, use of discriminative features and deep learning is marching towards new applications in security and surveillance, typically for tracking. Index Terms— Person Re-Identification, Camera Network, Video Analytics, Deep Learning, pedestrian detection. I. INTRODUCTION 1 With the spread of large networks of CCTV and surveillance cameras, it is beyond the capacities of a human operator to track an individual. Automated tracking systems are used to identify and tag a particular person by re- identifying. To identify the same human being across different frames of videos or still images is an important, but challenging task in intelligent video surveillance [1, 2]. The main objective of re-identification system is to find a person who appeared at instances and locations in the non- overlapping camera network. The basics of person re-identification abbreviated as re- ID are to compare a sought person or group(s) as seen in query image to a dataset or gallery of persons or group(s). If a person or group(s) in the query exist in the gallery, it results in a high matching ratio or higher index of similarity or rank compared to others. This involves inter person matching and intra-person matching both. Though person Re-ID is fundamentally similar to the image retrieval task, inter camera and intra-camera variances of viewpoint, illumination, occlusion, variations in pose, makes person Re-ID a considerably challenging issue to address [1]. Videos characteristically contain additional information than still images, video-based person re-identification has proved tougher with a new set of dimensions and challenges, turn out to be appealing for researchers [2]. Revised Manuscript Received on February 25, 2020. Manisha Talware, Research Scholar at G.H. Raisoni College of Engineering and Management, Pune, India Dr Sanjay Koli, Professor, D. Y. Patil Inst. of Info. Technology and Research Supervisor at G.H. Raisoni College of Engineering and Management, Pune, India Image-based person re-ids are generally categorized in feature representation and distance learning-based methods. Former aims at extracting distinctive features from pedestrian images like salience, midlevel, and salient colour features. The latter work to explore effective distance metrics, to increase matching accuracy based on the similarity index of two images. Recent methods include Large Margin Nearest Neighbor (LMNN), Relative Distance Comparison (RDC) and few others. Person re-id methods working on distance metrics learning are found effectiveness. However, these methods are basically evolved for image-based. These methods do not consider intricacies in video-based identification tasks. Person re-identification (re-ID) is one of the critical research area in video analytics, domains like security and surveillance in particular [2]. In real-life end-to-end surveillance arrangements, query image candidates are generated on-the-fly, resulting in dynamic gallery sets. As the surveillance video consists of numerous hours of videos and tons of person images, a scalable person Re-ID system with better efficiency along with generalization ability [1] is very much required to be designed. Apart from the volume, there exist large variation between different videos of pedestrian, even it is observed within each video, too. These variations make it a challenging problem to conduct re-ID between such pedestrian videos [3]. For significant improvement in the performance of Re-ID, generating robust descriptions [4], learning discriminative distance metrics [5], and powerful classifiers with good training is required. In practical, a video provides much information, so is a more natural method for person re-identification. Figure 1: Person Re-ID in Videos: End-to-End [6] Figure 1 illustrates the problem of person re-ID in videos. Severe deviations are observed between different sequences of the same individual [7] as reported in the iLIDS-VID dataset. [7] Large variations are also observed between different frames of the same video. The deviations among different videos of the same individual are termed as an inter-video variation. Variations in different frames of the same video are referred to as intra-video variations. Video-Based Person Re-Identification: Methods, Datasets, and Deep Learning Manisha Talware, Sanjay Koli