Abstract

Understanding public sentiment in crowded spaces has become essential for urban management, security monitoring, and event analysis. Traditional approaches often relied on surveys or manual observation, which are time-consuming and limited in scalability. Recent advancements in computer vision and artificial intelligence offered the potential for automated, real-time sentiment analysis. Monitoring emotions and behaviors in densely populated areas poses challenges such as occlusion, dynamic movement, and varying environmental conditions. Existing models often fail to achieve accurate detection in complex scenarios, limiting practical applications in safety, crowd management, and social analysis. This study employed a hybrid approach combining Generative AI techniques with the YOLO (You Only Look Once) object detection framework. YOLO was used to detect and track individual faces and body postures within the crowd. Generative AI was applied to enhance low-quality or partially occluded images and generate realistic feature representations for better emotion classification. Facial expressions, gestures, and body language were analyzed using a pre-trained sentiment recognition model. Data augmentation and feature normalization were applied to improve robustness and generalization. The proposed framework demonstrated significant improvements in detection and sentiment classification under dense and dynamic crowd conditions. Across multiple experiments, the system achieved an accuracy of 91.0%, precision of 89.1%, recall of 88.6%, F1-score of 89.0%, and MSE of 0.023, outperforming conventional Faster R-CNN, SSD-GAN, and Attention CNN-LSTM models by 6–12%. YOLO efficiently detected individual subjects, while generative enhancement minimized misclassification caused by occlusion and low-resolution inputs.

Authors

B. Yuvaraj, T. Ganesan, D.C. Jullie Josephine, S. Thumilvannan
Kings Engineering College, India

Keywords

Generative AI, YOLO, Sentiment Analysis, Crowd Monitoring, Real- Time Emotion Detection