Recently, there has been an increasing awareness of the critical role of sleep for the brain development of young infants. During the early neonatal stages of human development, the basic activity of the brain is to develop itself during sleep. Prolonged sleep is required in infants for further development of the nervous system. Sleep also plays an important role in body temperature regulation and energy saving. Neonatal sleep is divided into two main different sleep stages: Rapid EyeMovement (REM) and Non-Rapid Eye Movement (NREM). As the infant develops, sleep stages vary in maturity, length and distribution, thus underlining the importance of the quantification of these stages, which could eventually lead to new biomarkers of neonatal brain development. Nowadays, the golden standard in sleep monitoring is Polysomnography (PSG), in which vital signs as well as EEG and muscle activity are recorded during a whole-night study and subsequently sleep stages are classified by an expert. However, the high obtrusiveness of the multiple electrodes involved in PSG and its high associated costs make it impossible to use PSG as a routine monitor system. The SOMNUS project had two main targets: (i) to accurately measure respiration signals from patients using an ultra-wideband radar module, and (ii) to detect differences in respiration between REM and NREM phases in order to unobtrusively and automatically score sleep states of infants without the need of any electrode attached to the patient. The system has been developed using a training dataset of 23 patients from 3 months to 14 years old. It is for the first time UWB radar technology is used to monitor sleep in young patients. Moreover, this work provides a new data analysis algorithm to suppress motion artifacts from radar signals and increase the robustness of respiration monitoring. The results from the breathing detection algorithm developed in the present work provided an average mean absolute error of 3.25 Respirations per Minute in respiration rate. Amount of movement was employed to estimate Sleep/Awake events, with an average error in percentage of Total Sleep Time of 7.48%. Regarding sleep-state classification, several classification analyses were performed in order to study the best classification variable to detect REM/NREM, along with analyses regarding the best classification technique. Respiration variability was the main feature determining REM/NREMstate, with an overall sleep classification accuracy of around 80% when using linear classifiers such as Support Vector Machines and Fisher Linear Discrimination.