Post by AbdulRahman Zakaria

📱 AI Screen Time Monitoring: Fostering Healthier Digital Habits 📍 AIRE310 Project | Abu Dhabi University Excited to present our latest project: an AI-based system designed to help parents effectively monitor and manage their children's screen time. We built a solution utilizing deep learning to automatically detect both a child's presence and their screen interaction in a home environment. 🔍 Project Overview Our team developed parallel binary classifiers to process home video data: The Goal: Issue timely warnings to parents when a child's screen time exceeds healthy limits. Video Processing: We downsampled videos to 360p resolution at 4 frames per second using FFmpeg to optimize for efficiency. Dual Binary Classifiers: We implemented two classifiers—one to detect if a child is in the frame, and another to detect active screen interaction. CNN Architectures: We evaluated three mobile-friendly convolutional neural networks: EfficientNetV2, MobileNetV3, and MobileViT. 🧠 Tech Skills Applied This project heavily involved Computer Vision and Deep Learning pipelines: Data Preparation: Extracting frames, resizing, normalizing, and applying augmentation techniques like horizontal flipping and brightness adjustments. Deep Learning Framework: Training the models independently using PyTorch with GPU acceleration. Optimization: Applying binary cross-entropy loss, the Adam optimizer, and early stopping to prevent overfitting. System Integration: Combining outputs from both classifiers using logical operations to generate a final parental alert. 📊 Key Results EfficientNetV2 Superiority: EfficientNetV2 consistently outperformed the other architectures in both accuracy and AUC, demonstrating the best balance between model complexity and generalization on low-resolution frames. High Performance: All models showed exceptional capability in detecting child presence and screen interactions, yielding impressive accuracy, precision, and recall metrics. Robustness: The data augmentation strategies successfully improved the system's reliability across various room setups and lighting conditions. 📌 Lessons Learned Data Quality Matters: Model accuracy is deeply tied to the quality of the data, including consistent lighting and camera positioning. Hardware Constraints: Adapting to hardware limitations—such as choosing PyTorch over TensorFlow for stability—is a critical part of the engineering process. Ethical AI: Handling video data of children underscored the absolute necessity of privacy, consent, and ethical data processing. 👨‍💻 Team Members AbdulRahman Zakaria Abdullah Abouezz Anas Bassem Under the supervision of Dr. Yazan Dweiri #MachineLearning #ComputerVision #DeepLearning #ArtificialIntelligence #AIProjects #Engineering #AbuDhabiUniversity #CollegeOfEngineering