This paper presents a novel approach to solve data association in multi-camera multi-target object tracking. The main novelty is represented by the first known use of <i>dominant set framework</i> for intra-camera and inter-camera data association. Thanks to the properties of dominant sets, we can treat the data association as a global clustering of the detections (people or other targets) obtained over the whole sequence of frames from all the cameras. In order to handle occlusions, splitting and merging of targets, an efficient out-of-sample extension to dominant sets has been introduced to perform data association between different cameras (inter-camera data association). Experiments carried out on PETS '09 public dataset showed promising performance in terms of accuracy (precision and recall, as well as MOTA) when compared with the state of the art.