Multi-source homogeneous data clustering for multi-target detection from cluttered background with misdetection

This paper investigates a particular data mining problem which is to ‘identify’ an unknown number of targets based on homogeneous observations that are collected via multiple independent sources. This particular clustering problem corresponds to a significant problem of multi-target detection in the multi-sensor/scan context. No prior information is given about either the level of clutter (namely noisy data) or the number of targets/clusters, both of which have to be learned online from the data. In addition, the data-points from the same source cannot be grouped into the same cluster (namely the cannot link, CL, constraint) and the sizes of the generated clusters need to be bounded by the number of data sources. In the proposed approach, a density-based clustering mechanism is proposed firstly to identify dense regions as clusters and to remove clutter at the coarser level; the CL constraint is then applied for finer data mining and to distinguish overlapping clusters. Illustrative datasets are employed to demonstrate the validity of the present clustering approach for multi-target detection and estimation in cluttered environments which are affected by both misdetection and clutter.