A contribution to unobtrusive video-based measurement of respiratory signals
- Autores: Marc Mateu Mateus
- Directores de la Tesis: Juan Ramos Castro (dir. tes.), Mireya Fernández Chimeno (codir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2020
- Idioma: español
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Due to the growing popularity of video-based methods for physiological signal measurement, and taking into account the technological advancements of these type of devices, this work proposes a series of new novel methods to obtain the respiratory signal from a distance, based on video analysis. This thesis aims to improve the state of the art video methods for respiratory measurement, more specifically, by presenting methods that can be used to obtain respiratory variability or perform respiratory rhythm measurements. Moreover, this thesis also aims to present a new implementation of a time-frequency signal processing technique, to improve its computational efficiency when applied to the respiratory signals.
In this document a first approach to video-based methods for respiratory signal measurement is performed, to assert the feasibility of using a consumer-grade camera, not only to measure the mean respiratory rate or frequency, but to assert if this hardware could be used to acquire the raw respiratory signal and the respiratory rhythm as well. In this regard a new video-based method was introduced that measures the respiratory signal of a subject at a distance, with the aid of a custom pattern placed on the thorax of the subject.
Given the results from the first video-based method, a more broad approach was taken by comparing three different types of video hardware, with the aim to characterise if they could be used for respiratory signal acquisition and respiratory variability measurements. The comparative analysis was performed in terms of instantaneous frequency, as it allowed to characterise the methods in terms of respiratory variability and to compare them in the same terms with the reference method.
Subsequently, and due to the previous obtained results, a new method was proposed using a stereo depth camera with the aim to tackle the limitations of the previous study. The proposed method uses an hybrid architecture were the synchronized infrared frame and depth point-cloud from the same camera are acquired. The infrared frame is used to detect the movements of the subject inside the scene, and to recompute on demand a region of interest to obtain the respiratory signal from the depth point-cloud. Furthermore, in this study an opportunistic approach is taken in order to process all the obtained data, as it is also the aim of this study to verify if using a more realistic approach to respiratory signal analysis in real-life conditions, would influence the respiratory rhythm measurement.
Even though the depth camera method proved reliable in terms of respiratory rhythm measurement, the opportunistic approach relied on visual inspection of the obtained respiratory signal to properly define each piece. For this reason, a quality indicator had to be proposed that could objectively identify whenever a respiratory signal contained errors. Furthermore, from the idea to characterise the movements of a subject, and by changing the measuring point from a frontal to a lateral perspective to avoid most of the occlusions, a new method based on obtaining the movement of the thoraco-abdominal region using dense optical flow was proposed. This method makes us of the phase of the optical flow to obtain the respiratory signal of the subject, while using the modulus to compute a quality index.
Finally, regarding the different signal processing methods used in this thesis to obtain the instantaneous frequency, there were none that could perform in real-time, making the analysis of the respiratory variability not possible in real-life systems where the signals have to be processed in a sample by sample basis. For this reason, as a final chapter a new implementation of the synchrosqueezing transform for time-frequency analysis in real-time is proposed, with the aim to provide a new tool for non-contact methods to obtain the variability of the respiratory signal in real-time.
