In this paper, we present an algorithm for lane-level vehicle localization using lane region-segmented AVM images. First, we segment an input AVM image into lane regions. Then, we estimate a vehicle pose through ICP-based matching between the lane regions and the digital map. Contrary to previous methods that used restricted lane models, the proposed method facilitates vehicle localization even in arbitrary shaped road sections, as we utilize pixel-wise lane region itself without the use of the lane model. Through experiments in real roads, we verified the effectiveness of the proposed method.