A Multiple Layers Re-ranking Method for Person Re-identification

Pedestrian re-identification (re-ID) is a video surveillance technology for specific pedestrians in non-overlapping multi-camera scenes. However, due to the influence of dramatic changes in perspectives and pedestrian occasions, it is still a huge challenge to find a stable, reliable algorithm in high accuracy rate. In this paper, a multiple layers re-ranking approach is proposed to jointly account for that challenge. The re-ID is viewed as a multiple metrics ranking and optimizing problem by using a Multiple Layers Re-ranking framework. In this paper, multiple metrics are proposed by employing the correlation of different features to exploit comprehensive complementary information. Based on them, a multiple layers re-ranking framework is constructed to optimize and re-rank the initial results, which is more stable and effective than a single metric. Besides, a high similarity set is proposed to reduce the interference of appearance visual ambiguous samples. Through it, more effective candidate samples are selected into the re-ranking framework, improving the robustness. Experimental results on four person re-ID benchmarks demonstrate the effectiveness of the proposed framework. Especially, the matching rate of rank-1, our method outperforms the state-of-the-art methods on these datasets. Our code is available https://github.com/gengshuze/MLRL_re-id.git.