Occluded person re-identification based on feature fusion and sparse reconstruction

Abstract

Person Re-identification is one of the hotspots in the field of computer vision, especially for occluded person re-identification, which is still a challenge. In this paper, a feature fusion and sparse reconstruction based method of occluded person re-identification is proposed, which is suitable for person re-identification in various occlusion situations and where pose estimation is employed to obtain the occlusion body parts. A Full Occlusion Re-identification Network(FORN) is developed, where the obstruction is blackened. In the FORN, partial feature extraction and sparse feature reconstruction is combined through tree connections. The fusion features are facilitated in the FORN for occluded person similarity matching so that the matching rate of person re-identification under various occlusion situations is improved. On the occluded person re-identification datasets Partial-REID and Partial-iLIDS, the FORN method has obtained the experimental results of R-1 index 62.75% and 64.26%, and R-3 index 79.43% and 73.10%, respectively. Experiments are also conducted on conventional person re-identification datasets and the experimental results have verified the effectiveness and advancement of the proposed method.