Atrial activation (AA) detection during atrial fibrillation (AF) can be considered as a first step in estimating AF characteristics such as the AF cycle length. The detection of AA from intracardiac electrograms (EGM) remains challenging due to the constant variation in amplitude and duration of atrial EGM. This study is aimed at developing a robust detection of AA based on a novel non-linear filtering technique. Three consecutive patients (62-64 yrs.) with persistent AF (sustained AF duration 9-25 months) underwent catheter ablation (CA). Before CA, multipolar catheters were sequentially placed within the four pulmonary veins and the left atrial appendage for a duration of one minute. Sliding shortand long-term signal energies were measured for each sample in the EGM. A coefficient signal was then created as the ratio between the corresponding short- and long-term energies. Filtering was carried out by multiplying the coefficient signal to the EGM. Since AA have relatively higher amplitude than that of the noise and other EGM activities, the coefficient signal values are close to one where AA take place, while insignificant otherwise. Performance of the algorithm was measured with respect to activations manually annotated by a clinical expert. For a total of 5216 annotated activations, our method achieved a 99.6% detection rate, 99.8% specificity and 99.8% positive prediction value (PPV), against a state-of-theart approach [1] with 93.6%, 94.6% and 98.9% of detection rate, specificity and PPV respectively. These preliminary results indicate that the non-linear filter efficiently detects AA. This method is implementable by using two simple filters, avoids excessive use of arbitrary thresholds while incorporating physiological constraints. It offers low computational complexity, which makes it suitable approach for realtime/ online scenarios.