Published Paper Details:

: Personalized media experiences are rapidly evolving from static, preference-based models to dynamic, context-aware systems that respond in real-time to users' emotional states and activities. In this paper, we present a novel, integrated pipeline that fuses real-time facial emotion detection (captured via webcam) and offline activity recognition (analyzing uploaded video files) to drive a contextual song recommendation engine. The system comprises three tightly coupled modules: a Kivy-based GUI application leveraging OpenCV and DeepFace for low-latency facial affect analysis; a Flask web service for user management, video ingestion, and recommendation logic; and an offline video processor employing an Ultralytics YOLOv5 model fine-tuned for "running" and "sleeping" activities. We detail data collection and annotation procedures, model architectures and training regimes, algorithmic pseudocode, deployment via container orchestration, and front-end integration. Quantitative evaluation demonstrates 87-90% accuracy in seven-class emotion classification, 90.1% mAP in two-class activity detection, and round-trip latencies under 100 ms for emotion feedback. A user study with thirty participants reports 92% satisfaction with recommendation relevance and 4.6/5 mean perceived utility. Compared to standalone emotion- or activity-based recommenders, our unified approach yields a 25% uplift in personalization metrics. We conclude by mapping future research avenues: expanding affective and activity taxonomies, reinforcement-learning driven playlist adaptation, multimodal sensor fusion, and on-device inference for privacy.