Person Identification at a Distance
Abstract
This project aims to develop a software system that can take a surveillance video stream as input and automatically identify each person as either one specific known, or unknown, individual of the watch-list, when the person are at a far distance. Person identification can have tremendous applications to army, military, defense, and intelligence. For example, military might want to know, in real time, the identity of every individual in a certain area. Intelligence officers might search through a large volume of surveillance videos to find a few people who can match with the individuals in the watch-list. In both cases, person identification should be conducted at a distance as far as possible, which not only enables earlier identification thus allow ample time for action, but also provides more imagery observations over time for higher identification rates. Person identification is also a long-standing yet active topic in the research community. In the past decade, substantial research efforts have been made in face recognition, by working on constrained and unconstrained databases. The general consensus is that face recognition in the constrained and most of the unconstrained settings has been well solved. In comparison, the research of face recognition in the surveillance videos is only at its infancy, due to its unique challenges, including the low-resolution imagery, large pose discrepancy between the gallery and probe images. Therefore, given the interplaying among various challenges and the practical applications, face recognition in the surveillance videos is widely considered as the next big problem to be solved. In order to make strong impact to real-world applications and advance the state of the art for surveillance-video based face recognition, we propose a research project including three components. (1).We develop a joint face and person detection system, which would not only improve the detection rates for both face and person, but also enhance the alignment accuracy of the detected bounding boxes. (2) We develop a Generative Adversarial Network (GAN)-based algorithm to face image restoration. This algorithm can improve the spatial resolution of face images in surveillance videos, reduce the noise, and frontalize the face pose, while maintaining the identity of the faces. 3) We extend the GAN-algorithm to integrate the face image restoration with representation learning, where the objective is to learn a per-frame-based representation that is robust to resolution , quality, or pose and integrate these representations over the temporal domain. The ultimate goal of our project is to develop a suite of computer vision and deep learning algorithms, as well as an integrated software system, for searching a set of known individuals among surveillance videos, in either a live or a post-processing mode. The team is led by the computer vision and deep learning expert Dr. Xiaoming Liu at Michigan State University. The proposed project, if successful will transform the research on person identification at a distance that allows unprecedentedly vast quantities of surveillance videos to be exploited, leading to significant improvement in the recognition capability of person at a wider range of distances - a substantial contribution to the needs of Army. Defense, and Intelligence agencies. Specifically, our research will generate a software system with two fundamental capabilities. First, our proposed system will be able to achieve higher recognition rates of person identification, at a substantially far distance. Second, the ability to synthesize a higher-quality, restored, frontal-view and identity- preserved face image is also very valuable for soliciting witness for a specific suspect.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810330
Entities
People
- Xiaoming Liu
Organizations
- Army Contracting Command
- Michigan State University
- United States Army