The growing neural map: An on-line competitive clustering algorithm

On-line clustering is required whenever huge amounts or continuous flows of data must be classified and/or described in a compact way. Most available unsupervised competitive learning methods exhibit the problem that the units of the self-organizing map "accumulate" in the regions of the sample space characterized by higher data density, while in clustering it is generally desired that just one node represents each data cluster, independently of the relative density of points in the different clusters. This paper presents a novel on-line, hard-competitive algorithm that deals with this basic limitation, showing very good performance in clustering data from artificial distributions, as well as real data within the problem of motion-based scene segmentation in automated video surveillance.