Complex rhythmic patterns associated with Carnatic music are revealed from the stroke locations of percussion instruments. However, a comprehensive approach for the detection of these locations from composition items is lacking. This is a challenging problem since the melodic sounds (typically vocal and violin) generate soft-onset locations which result in a number of false alarms. In this work, a separation-driven onset detection approach is proposed. Percussive separation is performed using a Deep Recurrent Neural Network (DRNN) in the first stage. A single model is used to separate the percussive vs the non-percussive sounds using discriminative training and time-frequency masking. This is then followed by an onset detection stage based on group delay (GD) processing on the separated percussive track. The proposed approach is evaluated on a large dataset of live Carnatic music concert recordings and compared against percussive separation and onset detection baselines. The separation performance is significantly better than that of Harmonic- Percussive Separation (HPS) algorithm and onset detection performance is better than the state-of-the-art Convolutional Neural Network (CNN) based algorithm. The proposed approach has an absolute improvement of 18.4% compared with the detection algorithm applied directly on the composition items. © 2019 Jilt Sebastian, Hema A. Murthy.