Remote photoplethysmography (rPPG) has gained increasing importance, especially during and after the COVID-19 pandemic, for its ability to estimate heart rate by analyzing subtle variations in skin color without physical contact. This non-invasive and practical method relies on capturing changes in pixel intensity through RGB video or near-infrared imaging. In this study, we propose a novel hybrid model that leverages the feed-forward integration of 3D Convolutional Neural Networks (3D-CNNs) and Video Vision Transformers (ViViTs) to enhance rPPG estimation. The 3D-CNNs first capture local spatiotemporal features, using Temporal Central Difference Convolution (3DCDC-T) and Convolutional Block Attention Module (CBAM). These local features are then passed to the Video Vision Transformer, where Multi-Head Self-Attention (MHSA) captures global contextual relationships and long-range dependencies across frames, enabling more effective representation of complex temporal dynamics. This sequential learning allows the model to progressively refine features from local to global, ensuring more consistent and coherent feature extraction. Our feed-forward approach also improves computational efficiency by reducing the dimensionality of the input data before global attention processing, making it particularly effective in datalimited environments. Through comprehensive experiments, we show that our hybrid approach outperforms state-of-the-art methods across multiple public datasets, achieving a 22.55% improvement in MAE and a 55.80% improvement in RMSE on the UBFC-rPPG dataset, demonstrating superior feature progression and generalization in rPPG and heart rate estimation tasks.

Non-contact vital sign monitoring in Pediatric Intensive Care Units is challenged by frequent occlusions, data scarcity, and the need for temporally stable anatomical tracking to extract reliable physiological signals. Traditional detectors produce unstable tracking, while video transformers are too computationally intensive for deployment on resource-limited clinical hardware. We introduce Divided Space–Time Mamba, an architecture that decouples spatial and temporal feature learning using State Space Models to achieve linear-time complexity, over 92% lower than standard transformers. To handle data scarcity, we employ self-supervised pre-training with masked autoencoders on over 50 k domain-specific video clips and further enhance robustness with multimodal RGB-D input. Our model demonstrates superior performance, achieving 0.96 mAP@0.5, 0.62 mAP50-95, and 0.95 rotated IoU. Operating at 23 FPS (43 ms latency), our method is approximately 1.9× faster than VideoMAE and 5.7× faster than frame-wise YOLOv8, demonstrating its suitability for real-time clinical monitoring.