The omnipresence of using Enterprise Resource Planning (ERP) systems to support business processes has enabled recording a great amount of (relational) data which contains information about the behaviors of these processes. Various process mining techniques have been proposed to analyze recorded information about process executions. However, classic process mining techniques generally require a linear event log as input and not a multi-dimensional relational database used by ERP systems. Much research has been conducted into converting a relational data source into an event log. Most conversion approaches found in literature usually assume a clear notion of a case and a unique case identifier in an isolated process. This assumption does not hold in ERP systems where processes comprise the life-cycles of various interrelated data objects, instead of a single process. In this paper, a new semi-automatic approach is presented to discover from the plain database of an ERP system the various objects supporting the system. More precisely, we identify an artifact-centric process model describing the system's objects, their life-cycles, and detailed information about how the various objects synchronize along their life-cycles, called interactions. In addition, our artifact-centric approach helps to eliminate ambiguous dependencies in discovered models caused by the data divergence and convergence problems and to identify the exact "abnormal flows". The presented approach is implemented and evaluated on two processes of ERP systems through case studies.