To precisely ablate tumor in radiation therapy, it is important to locate the tumor position in real time during treatment. However, respiration-induced tumor motions are difficult to track. They are semi-periodic and exhibit variations in baseline, frequency and fundamental pattern (oscillatory amplitude and shape). In this study, we try to decompose the above-mentioned components from discrete observations in real time. Baseline drift, frequency (equivalently phase) variation and fundamental pattern change characterize different aspects of respiratory motion and have distinctive clinical indications. Furthermore, smoothness is a valid assumption for each one of these components in their own spaces, and facilitates effective extrapolation for the purpose of estimation and prediction. We call this process 'profiling' to reflect the integration of information extraction, decomposition, processing and recovery. The proposed method has three major ingredients: (1) real-time baseline and phase estimation based on elliptical shape tracking in augmented state space and Poincaré sectioning principle; (2) estimation of the fundamental pattern by unwarping the observation with phase estimate from the previous step; (3) filtering of individual components and assembly in the original temporal-displacement signal space. We tested the proposed method with both simulated and clinical data. For the purpose of prediction, the results are comparable to what one would expect from a human operator. The proposed approach is fully unsupervised and data driven, making it ideal for applications requiring economy, efficiency and flexibility.