Age Detection in a Surveillance Video Using Deep Learning Technique

• Published in: SN computer science 2021-07, Vol.2 (4), p.249, Article 249
• Main Authors: Vasavi, S., Vineela, P., Raman, S. Venkat
• Format: Article
• Language: English
• Subjects: Accuracy; Age; Algorithms; Artificial intelligence; Biometrics; Classification; Closed circuit television; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Correlation coefficients; Data Science and Communication; Data Structures and Information Theory; Datasets; Deep learning; Face recognition; Feature extraction; Forensic computing; Gender; Image processing; Image reconstruction; Information Systems and Communication Service; Literature reviews; Machine learning; Neural networks; Object recognition; Original Research; Pattern Recognition and Graphics; Social networks; Software Engineering/Programming and Operating Systems; Surveillance; Video data; Vision; Workflow
• ISSN: 2662-995X; 2661-8907
• DOI: 10.1007/s42979-021-00620-w

The video streams that are collected from CCTV surveillance camera can be used in many applications such as crowd analysis, forensic, self-profile analysis, and social network user’s analysis. Soft biometrics such as age, gender, height, skin color can be used for human analysis. This requires object detection and feature analysis. Works reported for face recognition and age detection have poor performance with real-world profile images. It may be because of incomplete description of the human object. Also, they rely on traditional image processing algorithms that extract hand-crafted features. Deep learning workflow transforms the identified patterns into mathematical modeling that can be used for subsequent prediction. Residual networks can skip connections and can address vanishing gradient problem with improved accuracy. Wide ResNet 34-based system is proposed in this paper that automatically predicts age of human object in video images. Modified Wide ResNet is used for feature extraction that learns facial keypoints, image reconstruction using Simultaneous algebraic reconstruction technique for up sampling, 101 number of classes (101-way classification) ranging from 0 to 100. Proposed system accuracy is evaluated using mean absolute error and with Pearson correlation coefficient that finds correlation between actual age and predicted age. Experimental results proved that data augmented Wide ResNet out performs the existing age prediction methods with 5% increase in accuracy.