基于跨场景迁移学习的行人再识别

To solve the problem of low recognition rate of the pedestrian re-identification system in new scenarios, a pedestrian re-identification method based on cross-scenario transfer learning is proposed. Due to the scarcity of labeled information in new scenarios, pedestrian images obtained in other scenarios are utilized to assist in the pedestrian re-identification of the target scenario. The images are preprocessed with Retinex transformation to reduce the impact of illumination. The datasets of different scenarios are jointly learned through an asymmetric multi-task learning method to obtain a similarity measurement function. A cross-task data difference model is employed to address the data collapse problem in the shared space, and a regularization term is introduced to alleviate the overfitting phenomenon. The experimental results demonstrate that, compared with relevant pedestrian re-identification methods without transfer learning, the cross-scenario transfer learning method exhibits a significant improvement in the recognition rate.

为解决在新场景中行人再识别系统识别率低的问题，提出基于跨场景迁移学习的行人再识别方法。由于新场景中已标记信息很少，利用在其它场景中获得的行人图像来 帮 助 目 标 场 景 的 行 人 再 识 别。对图像用Retinex变换预处理，减少光照影响，通过非对称多任务学习方式联合学习不同场景的数据集，得到相似性度量函数，运用跨任务数据差异模型解决共享空间的数据塌陷问题，引入正则化项，缓解过拟合现象。实验结果表明，相比有关非迁移学习的行人再识别方法，基于跨场景迁移学习方法在识别率上有很好的提升。