Electrocardiogram (ECG) is the measurement of potential differences occurring on the body due to the currents that flow on the heart during diastole and systole. Cardiac abnormalities cause uncommon current flows, leading to strange waveform morphologies in the recorded ECG. Since some abnormalities become visible in ECG only during activity, exercise ECG tests are conducted. The sources of noise during an exercise test are electro myogram (EMG) due to increased muscle activity and baseline wander (BW) due to mechanical motion. Frequency band filtering, used to eliminate noise, is not an efficient method for filtering noise because usually frequency spectra of the interference and the ECG overlap. Rather, a fast morphological filter is required. This thesis is focused on an online filtering approach which separates noise and ECG signals without changing the morphology. The redundancy present in standard 12 lead ECG records is made operational by a Singular Value Decomposition based orthogonalization of the input signals. ECG is represented in a minimum dimensional space whose orthogonal complement takes on noise. The signals in this low dimensional space are used to reconstruct the input signals without noise. Noise elimination also improves data compression. A comparative study of the ST analysis of original and reconstructed signals is presented at the end.