基于生成对抗网络联合时空模型的行人重识别方法

跨摄像头的行人跟踪是智慧城市和智能安防的技术难题，而行人重识别是跨摄像头行人跟踪中最关键的技术。针对因领域偏差而导致行人重识别算法跨场景应用时准确率大幅下降的问题，提出一种基于生成对抗网络（GAN）联合时空模型的方法（STUGAN）。首先，通过引入GAN生成目标场景的训练样本，以增强识别模型的稳定性；然后，利用时空特征构建目标场景样本的时空模型，以筛选低概率匹配样本；最后，融合识别模型与时空模型来进行行人的重识别。在该领域典型的公开数据集Market-1501和DukeMTMC-reID上与词袋模型（BoW）、先进的无监督学习（PUL）、无监督多任务字典学习（UMDL）等先进无监督算法进行实验对比。实验结果分析表明，所提方法在Market-1501数据集上的rank-1、rank-5及rank-10指标的准确率达到66.4%、78.9%及84.7%，比对比算法最好结果分别高5.7，5.0和4.4个百分点；平均精度均值（mAP）仅低于保持主题一致性的循环生成网络（SPGAN）。

Person re-identification method based on GAN uniting with spatial-temporal pattern

Tracking of the person crossing the cameras is a technical challenge for smart city and intelligent security. And person re-identification is the most important technology for cross-camera person tracking. Due to the domain bias, applying person re-identification algorithms for cross-scenario application leads to the dramatic accuracy reduction. To address this challenge, a method based on Generative Adversarial Network (GAN) Uniting with Spatial-Temporal pattern (STUGAN) was proposed. First, training samples of the target scenario generated by the GAN were introduced to enhance the stability of the recognition model. Second, the spatio-temporal features were used to construct the spatio-temporal pattern of the target scenario, so as to screen low-probability matching samples. Finally, the recognition model and the spatio-temporal pattern were combined to realize the person re-identification task. On classic datasets of this field named Market-1501 and DukeMTMC-reID, the proposed method was compared with BoW (Bag-of-Words), PUL (Progressive Unsupervised Learning), UMDL (Unsupervised Multi-task Dictionary Learning) and other advanced unsupervised algorithms. The experimental results show that the proposed method achieves 66.4%, 78.9% and 84.7% recognition accuracy for rank-1, rank-5 and rank-10 indicators on the Market-1501 dataset respectively, which are 5.7, 5.0 and 4.4 percentage points higher than the best results of the comparison algorithm, respectively; and the mean Average Precision (mAP) higher than the comparison algorithms except Similarity Preserving cycle-consistent Generative Adversarial Network (SPGAN).