The overall aim of the research reported in this thesis was to build a new non-invasive method for Heart Rate monitoring using videos and investigate the algorithms for selecting a region of interest to enable the extractions of more useful features for the measurement of Heart Rate. Analysis of these image features will assist in the creation of a more effective database for monitoring Heart Rate.

Heart Rate is important for monitoring and diagnosing medical conditions of a individuals health. Several techniques are available to measure heart rate, these used sensors, which need to be in direct contact with the surface of the skin, that may cause discomfort and soreness to the patient, especially for sensitive skin patients. However, recent advances in computer vision have shown that several physiological factors linked to the heart may be evaluated without invasive methods. Heart rate monitoring can be done using video observation on the subject by analysing specific facial parts like eyes, forehead, cheeks etc. The literature has observed and shown that the facial colour changes, leading to redness from the normal face colour when the heart rate goes up. However, the changes are so minute that they cannot be observed from the naked eye, the frame responsible for detecting needs to be magnified. This research work has utilised Euler video modification as it has been cited as one of the most efficient techniques for magnifying specified portions of the face. The research was carried out by recording 30 participants, and a pulse oximeter was used to measure the heart rate.

The research first extracts the frames from the input video, and then face detection is done using cascaded object detection. Then, evaluate the face's forehead using histogram equalization, magnifies it, and optimize the magnified part using Swarm Intelligence oriented Artificial Bee Colony (ABC). Modified grouping behaviour has been presented, and a novel fitness function based on the pixel distribution has been applied. Due to the shortage of samples for the training section as the global pandemic affected the entire world, the training has been done by conventional neural networks that work with Levenberg Back Propagation algorithm.

The result section has been designed based on frame analysis with true-positive rate, false-positive rate, Accuracy and Kappa coefficient. The results contain the analysis based on the variation of the total number of samples in the data repository. The results compared analysis of the algorithms used in this research, Convolutional Neural Network (CNN) and Feed Forward Back Propagation (FFBPNN) with optimization method ABC and without optimization. Overall, the results compared with the proposed algorithm FFBPNN-L with ABC compared to CNN with ABC that TPR has been improved by 8.24%, Kappa coefficient has been improved by 4.58%. The limited dataset shows better results for FFBPNN-L, such that ABC with FFBPNN-L resulted in 4.25% better accuracy than ABC with CNN. Thus, FFBPNN-L with the integration of ABC shows improved results than when ABC is integrated with CNN. CNN extracts the features and does not require any external feature extraction support. This outcome with better performance of FFBPNN over CNN is mainly due to the smaller dataset size available for the research.

The findings of this research, which demonstrate methodologies that can be used to monitor heart rate, also identify directions for future work in medical engineering