多特征融合与独立测度学习的行人再识别

目的 由于行人图像受到光照、视角、遮挡和行人姿态等变化的影响，在视觉上容易形成很大的外观差异，对行人再识别造成干扰。为了提高行人再识别的准确性，针对以上问题，提出一种基于多特征融合与独立测度学习的行人再识别算法。方法 首先通过图像增强算法对原始图像进行处理，减少因光照变化产生的影响，然后对处理后的图像进行非均匀分割，同时提取行人图像的HSV、RGS、LAB和YCbCr 4种颜色特征和SILTP（scale invariant local ternary pattern）纹理特征，在基于独立距离测度学习方法下，融合行人的多种特征，学习得到行人图像对的相似度度量函数，最后将行人图像对的相似度进行加权匹配，实现行人再识别。结果 在VIPeR、iLIDS和CUHK01这3个数据集上进行实验，其中Rank1（排名第1的搜索结果即为待查询人的比率）分别达到42.7%、43.6%和43.7%，Rank5（排名前5的搜索结果中包含待查询人的比率）均超过70%，识别率有了显著提高，具有实际应用价值。结论 提出的多特征融合与独立测度学习的行人再识别算法，能够有效表达行人图像信息，且对环境变化具有较强的鲁棒性，有效提高了识别率。

Person re-identification based on multi-features fusion and independent metric learning

Objective Person re-identification is a very challenging problem and has practical application value. It plays an important role in video surveillance systems because it can reduce human efforts in searching for a target from a large number of videos. However, the pedestrian''s image is easily affected by illumination changes, different viewpoints, varying poses, complicated background and the problem of occlusion and scale. It is likely to form a lot of differences in appearance and that causes interference in person re-identification. To solve this problem, many studies concentrate on designing a feature representation or metric learning method. For the above problem, this study proposes a robustness algorithm based on multi-features fusion and independent metric learning for person re-identification. Method First, the original images are processed by image enhancement algorithm to reduce the impact of illumination changes. This enhancement algorithm is committed to making the image closer to the human visual characteristics. Then, using the method of non-uniform segmentation processes images. At the same time, processed images are extracted from four color features including HSV, RGS, LAB and YCbCr feature and a texture feature of SILTP (scale invariant local ternary pattern).What''s more, through multi-features fusion and independent metric learning, the algorithm gets a similarity measure function of the related person. Finally, the algorithm weights the original similarity and gets the ultimate similarity achieving person re-identification. Result The proposed method is demonstrated on three public benchmark datasets including VIPeR, iLIDS and CHUK01. Each dataset has its own different characteristics. And experimental results show that the proposed method achieves a higher accuracy rate with excellent features and particular method of fusion and learning compared with other similar algorithms. The proposed method achieves a 42.7% rank-1 (represents the correct matched pair) on VIPeR benchmark and respectively 43.6% and 43.7% on iLIDS and CHUK01 benchmark. It is worth mentioning that the rank-5 (represents the expectation of the matches at rank 5) are more than 70% on the three datasets. It greatly improves the recognition rate and has practical application value. Conclusion The experimental results show that the proposed method can more effectively express pedestrian''s image information. Furthermore, the proposed method has strong robustness to illumination changes, different viewpoints, varying poses, complicated background and the problem of occlusion and scale and effectively improves the recognition rate.