GAIT recognition using silhouette using CNN, Transfer Learning

Abstract

Background: GAIT means walking pattern. This walking pattern can be used to authenticate person in public areas. This biometric recognition system can identify persons even in low resolution images captured from surveillance camera.
Methods: Paper outlines feasibility in identification of person using GAIT as a biometric security system. Also briefs importance of contactless biometric based security systems are formulated considering videos of person (marker-free GAIT). Paper addresses designing of custom Convolutional Neural Network (CNN) and compare performance transfer learning models. Person identification is performed using silhouette captured indoor & outdoor.
Results: Models employed are Deep Learning method CNN, three transfer learning (VGG16, Xception and MobileNetV2). Result analysis shows that CNN achieved 93% for indoor and 91% for outdoor. Proposed model is tested on CASIA-B standard benchmark’s silhouette and achieved accuracy of 87%. Hence propose model is efficient in detecting persons for real time applications. It is implemented using python scripting language.