| The problem of person re-identification is to identify and associate people when they appear at different times and physical locations.In recent years,person re-identification has attracted more and more attention in academic research and industrial experiment.However,the task of person re-identification is still facing many challenges,such as the robustness of feature description,small sample size,poor generalization performance and so on.In view of these challenges,this thesis proposes an algorithm integrating traditional features and depth features based on AdaRank.There are four main innovations in the algorithm:1.The original HOG3D feature is modified to adapt to practical datasets.2.The traditional LOMO feature is upgraded to LOM03D feature so that it can capture the time information effectively.3.Using 3D convolutional neural networks instead of 2D convolutional neural networks.In this process,the training samples are selectively used and the two metric methods are used to improve the training efficiency.4.Using AdaRank to integrate the resulting weak rankers to improve the overall performance of the system.Through comparison experiments,the three main innovation points of this algorithm contribute to the overall performance,and the overall performance of the algorithm is superior to state-of-the-art algorithms. |
