DBSCAN, a fundamental density-based clustering method, is well-known for its ability to discern clusters of diverse shapes. However, its effectiveness diminishes with the scale of datasets, as the storage and processing of labels become impractical. Conversely, centroid-based algorithms like K-means adeptly handle raw data by assigning them to the nearest centroids, yet struggle with non-spherical clusters. This study aims to amalgamate the strengths of these distinct clustering methodologies. Proposing the Incremental Prototype-based DBSCAN (IPD) algorithm, our study addresses the challenge of identifying clusters with arbitrary shapes within large datasets. IPD not only identifies clusters but also selects representatives for each, with a primary focus on establishing a stability criterion within the prototype space. Experimental results demonstrate the efficacy of the approach, showcasing perfect convergence and scalability, even with extremely large synthetic datasets containing shaped clusters. The code for the IPD algorithm is available at https://github.com/Jayasree-Saha/IPD .