Multimodal Analysis in Short-Form Video Content: Integrating Skin Exposure Detection and Body Posture Recognition on TikTok

This thesis introduces a computational approach for automated skin exposure analysis in TikTok videos using a multi-stage computer vision pipeline. The method involves frame extraction, posture analysis, body part segmentation, and skin exposure quantification. A custom CNN model was trained to classify skin exposure across different body regions. The system was tested on 5,000 TikTok videos with diverse demographics and lighting conditions, achieving 93.78% accuracy. Results show effectiveness in handling skin tones and complex backgrounds. This research enhances content moderation tools and addresses challenges in automated visual content analysis.