Novel Similarity Metric Learning Using Deep Learning and Root SIFT for Person Re-identification
- PDF / 1,703,290 Bytes
- 17 Pages / 439.37 x 666.142 pts Page_size
- 12 Downloads / 232 Views
Novel Similarity Metric Learning Using Deep Learning and Root SIFT for Person Re‑identification M. K. Vidhyalakshmi1 · E. Poovammal1 · Vidhyacharan Bhaskar2 · J. Sathyanarayanan3 Accepted: 29 October 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract This paper deals with the person re-identification, intending to match the images of the person captured using disjoint cameras mounted in different locations. Such a task of matching the images remains a difficult issue as the appearance of the individual differs from the perspective of the various cameras. Inspired by the recent success of deep learning in the domain of person re-identification, a novel deep learning framework which combines deep features and Root Scale Invariant Features Transform (Root SIFT) features has been proposed. The conventional deep Convolutional Neural Network (CNN) can obtain significant features but does not take into account the spatial relationship between the features. Also, CNN requires an enormous number of instances to train the network. To address these issues, the proposed method combines Root SIFT features along with the CNN features. With the combination of Deep and Root SIFT features, the model can give improved performance over other CNN based models. Experiments were conducted on standard datasets CUHK 03 (labelled and detected), CUHK 01 and VIPeR and the matching rate is reported as 74.45% for CUHK 03 (labelled), 72.63% for CUHK 03 (detected), 76.12% for CUHK 01 and 48.45% for VIPeR dataset. The experiments demonstrate that the proposed algorithm has improved identification rate over the recent algorithms. Keywords Person re-identification · CNN · Siamese network · Root SIFT · Bag of visual words
1 Introduction In many places, camera monitoring systems have been installed for surveillance purpose. These monitoring systems have a set of cameras fixed at different locations. Such images of a person obtained from these cameras vary considerably. Because these images are taken by cameras under different lighting conditions, various pose changes, viewpoint variation, * Vidhyacharan Bhaskar [email protected] 1
Department of CSE, SRMIST, Kattangulathur, Chennai, India
2
Department of Electrical and Computer Engineering, San Francisco State University, San Francisco, CA, USA
3
R & D, Dev Technologies, Chennai, India
13
Vol.:(0123456789)
M. K. Vidhyalakshmi et al.
occlusions and low resolution; it is hard to associate the image of a person taken by one camera with the other. The problem has attracted a lot of researchers’ attention to explore and find the best solution, this emerging problem is called person re-identification. Reidentification is a process of associating a person’s images taken by different cameras positioned at different locations. Challenges to this issue include lighting adjustments, lowgrade sensors, different camera settings and occlusions. The established methods initially concentrated on extracting the hand-crafted invariant features from the images and use me
Data Loading...
