Probabilistic matching pair selection for SURF-based person re-identification

Abstract

The objective of this paper is to study the performance of human reidentification based on multi-shot SURF and to assess its degradation according to the angular difference between the test and reference video scene view angles. In this context, we propose a new automatic statistical method of acceptance and rejection of SURF correspondence based on the likelihood ratio of two GMMs learned on the reference set and modeling the distribution of distances resulting from matching sequences associated with the same person and with different persons respectively. The experimental results show that our approach compares favorably with the state of the art and achieves a good performance.