A Hierarchical Frame-by-Frame Association Method Based on Graph Matching for Multi-object Tracking

Abstract

Multiple object tracking is a challenging problem because of issues like background clutter, camera motion, partial or full occlusions, change in object pose and appearance etc. Most of the existing algorithms use local and/or global association based optimization between the detections and trackers to find correct object IDs. We propose a hierarchical frame-by-frame association method that exploits a spatial layout consistency and inter-object relationship to resolve object identities across frames. The spatial layout consistency based association is used as the first hierarchical step to identify easy targets. This is done by finding a MRF-MAP solution for a probabilistic graphical model using a minimum spanning tree over the object locations and finding an exact inference in polynomial time using belief propagation. For difficult targets, which can not be resolved in the first step, a relative motion model is used to predict the state of occlusion for each target. This along with the information about immediate neighbors of the target in the group is used to resolve the identities of the objects which are occluded either by other objects or by the background. The unassociated difficult targets are finally resolved according to the state of the object along with template matching based on SURF correspondences. Experimentations on benchmark datasets have shown the superiority of our proposal compared to a greedy approach and is found to be competitive compared to state-of-the-art methods. The proposed concept of association is generic in nature and can be easily employed with other multi-object tracking algorithms.