Abstract:

Person re-identification is an important topic in the fields of computer vision and public safety systems based on video surveillance. With the development of large-scale data sets and extensive research on deep learning networks, deep learning has achieved good results in person re-identification problems. However, in addition to different visual appearance changes caused by different cameras and different perspectives of the video itself, the overall data set for person re-identification is small, and the number of labeled training data samples is smaller, which limit the person re-identification effect of deep learning. An improved deep transfer learning algorithm for person re-identification is proposed in this paper, which migrates mature deep learning models that have been trained on large data sets to the target data set for fine-tuning, and combines the target data to optimize the model to make it better feature extraction for a new data set. During the model training process, by improving the triplet loss function, the distance between the same samples is closer, and the distance between different samples is farther. At the same time, the distance between positive samples is set to be not greater than a threshold, which ensures that the clusters formed in the feature space will not be too large. The algorithm in this paper reduces the time for training deep learning models, avoids the shortcomings of insufficient data on small data sets, and improves the accuracy of person re-identification. Comparative experiments on five standard data sets show that the proposed algorithm has achieved good person re-identification results.

Key words: person re-identification, deep learning, transfer learning, triplet loss

摘要：

跨镜行人追踪是计算机视觉和视频监控公共安全体系构建等领域的重要课题。伴随大规模数据集的发展和深度学习网络的广泛研究，深度学习在跨镜行人追踪问题中取得了良好效果。然而在应用中，除了监控视频自身的不同摄像头、不同视角引起的不同视觉表象变化外，面向跨镜行人追踪的整体数据集偏小，具有标记的训练数据样本量更小，从而制约了基于深度学习的跨镜行人追踪效果。提出了改进型深度迁移学习的跨镜行人追踪算法，将在大数据集上训练好的成熟模型进行微调并迁移到目标数据集上，结合目标数据进行优化，使其能更好地针对新数据集做特征提取。在模型训练过程中，通过改进三元组损失函数，拉近相同样本之间的距离，加大不同样本之间的距离，同时设定正样本之间的最大距离阈值，从而保证特征空间生成的簇不会太大，利于模型的优化。该算法减少了深度学习训练模型的时间，避免了小数据集上数据量不足等缺点，提高了跨镜行人追踪的准确度。在五个基准数据集上的跨镜行人追踪对比实验显示，改进算法取得了良好效果。

关键词: 跨镜行人追踪, 深度学习, 迁移学习, 三元组损失