Region of Interests (ROIs) generation plays a critical role in pedestrian detection systems. The challenge lies in generating as few ROIs as possible at low computational complexity, while ensuring none of the pedestrians in the scene are omitted. However, existing ROIs generation methods either result in a large number of irrelevant ROIs or are compute-intensive. In addition, distinguishing pedestrians who are in close proximity is still a big challenge. In this paper, we propose an efficient stereo based ROIs generation method that is based on a two-level incremental segmentation strategy. An adaptive strategy is first employed to identify a minimal set of clusters from the u-disparity image. The initial clusters are then further refined to distinguish pedestrians in close proximity. Experimental results based on a challenging benchmark show that the proposed algorithm outperforms two state-of-art baseline algorithms by being able to distinguish pedestrians in close proximity with a small number of ROIs.