Development of an Intelligent System for User Segmentation in Crowded Environments Using Facial Recognition an‎d Motion Analysis Based on Deep Neural Networks an‎d Multidimensional Data Analysis

10/19/2026 12:00:00 AM

Person identification in crowded environments is one of the significant challenges in the field of computer vision an‎d deep learning. This research aims to develop an intelligent multimodal system for accurate user recognition under complex environmental conditions. The proposed approach is based on combining two biometric features: face an‎d gait (walking pattern), which have the capability to compensate for each other’s limitations. In this research, two face recognition models, FaceNet an‎d AdaFace, were first implemented an‎d eva‎luated. The FaceNet model with InceptionResnetV1 architecture, using the LFW dataset containing 60,234 images from 5,749 individuals, achieved an accuracy of 95.66%. The AdaFace model with ResNet-100 architecture an‎d adaptive loss function demonstrated superior performance with an accuracy of 99.82%. For gait recognition, the EnhancedGaitCNN model with four convolutional blocks was developed an‎d trained using the CASIA-B dataset containing 124 subjects under various conditions an‎d viewing angles. Subsequently, two fusion models comprising the combination of FaceNet with EnhancedGaitCNN an‎d the combination of AdaFace with EnhancedGaitCNN were implemented using various score-level fusion methods. The fusion methods included fixed weighting, rule-based approaches, quality-based methods, an‎d machine learning techniques. Results indicated that the fusion model of AdaFace an‎d EnhancedGaitCNN with weights of 0.85 for face an‎d 0.15 for gait provides the best performance. eva‎luation of the fusion models on the SMVDU-Single-Gait dataset showed that the combined FaceNet an‎d EnhancedGaitCNN model with fixed weight method of 0.85 achieved a rank-1 accuracy of 97.01%. The combined AdaFace an‎d EnhancedGaitCNN model with the same weighting method demonstrated superior performance with a rank-1 accuracy of 98.50%, which represents a significant improvement compared to using either model individually. These results confirm the system’s applicability in surveillance systems, access control, an‎d person identification in real-world environments.

تشخيص چهره , تشخيص گيت , تشخيص كاربر , يادگيري عميق , ادغام چندوجهي

Face Recognition , Gait Recognition , User Identification , Deep Learning , Multimodal Fusion

Parsa Asasi Moghaddam

Dr. Majid Rajabi