This paper presents a comprehensive approach to stress detection using remote photoplethysmography (rPPG) and deep learning techniques. The study focuses on detecting stress through facial videos, utilizing hybrid deep learning (DL) networks based on rPPG signals. The proposed method employs Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and 1D Convolutional Neural Network (1D-CNN) models, with hyperparameter optimization and data augmentation to enhance performance. The system achieves up to 95.83% accuracy in stress detection using the UBFC-Phys dataset, demonstrating high efficiency and effectiveness. The methodology involves data preprocessing, including Fast Fourier Transform (FFT) for frequency domain analysis and data augmentation techniques like linear interpolation and Gaussian white noise. The study evaluates the performance of different DL models on both ground-truth (GT) BVP signals and rPPG signals, showing that the 1D-CNNv1 model achieves the highest accuracy of 95.83% when using the CuPy-CHROM method and white noise augmentation. The results indicate that the proposed hybrid DL models significantly improve stress detection accuracy and efficiency, making them suitable for real-time applications. The study also highlights the importance of privacy considerations in rPPG-based stress detection due to the use of cameras and the diversity of participants. Future work will focus on improving signal extraction through alternative physiological sensing tools and optimizing existing toolboxes. The study contributes to the field of stress detection by introducing a novel approach that combines non-contact and physiological techniques for continuous monitoring of biomedical signals.This paper presents a comprehensive approach to stress detection using remote photoplethysmography (rPPG) and deep learning techniques. The study focuses on detecting stress through facial videos, utilizing hybrid deep learning (DL) networks based on rPPG signals. The proposed method employs Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and 1D Convolutional Neural Network (1D-CNN) models, with hyperparameter optimization and data augmentation to enhance performance. The system achieves up to 95.83% accuracy in stress detection using the UBFC-Phys dataset, demonstrating high efficiency and effectiveness. The methodology involves data preprocessing, including Fast Fourier Transform (FFT) for frequency domain analysis and data augmentation techniques like linear interpolation and Gaussian white noise. The study evaluates the performance of different DL models on both ground-truth (GT) BVP signals and rPPG signals, showing that the 1D-CNNv1 model achieves the highest accuracy of 95.83% when using the CuPy-CHROM method and white noise augmentation. The results indicate that the proposed hybrid DL models significantly improve stress detection accuracy and efficiency, making them suitable for real-time applications. The study also highlights the importance of privacy considerations in rPPG-based stress detection due to the use of cameras and the diversity of participants. Future work will focus on improving signal extraction through alternative physiological sensing tools and optimizing existing toolboxes. The study contributes to the field of stress detection by introducing a novel approach that combines non-contact and physiological techniques for continuous monitoring of biomedical signals.