最後為了克服2D卷積網路難以取得區域時序特徵的問題，本論文提出了一個方法在網路中引入部分3D卷積結構，可以有效利用有限的記憶體資源設法取得更多有效的特徵，使得網路能夠發揮更佳的性能。 ;The topic of this thesis is the implementation and improvement of convolutional neural networks applied to gait recognition methods. The purpose is training a convolutional neural network to extract the gait features of the human walking sequence, which preprocessed by detecting and cutting the ROI of person in the RGB image sequence. The extracted features extracted will use to identify people.
Gait recognition is a non-contact biometric method to determine the identity or physical condition of people by analyzing the different postures and habits of people performing when they are walking, including skeletons and joint movements.
Different from the method based on the detection skeleton, this paper uses the pre-training model(YOLOv2 and FlowNet2.0) to extract the optical flow feature maps and ROIs from the input sequence. Cutting and Concatenating the optical flow feature maps as the low-level feature input. Then, we will train a model built with the Wide Residual Network architecture to extract high-level abstract features from the low-level feature. And we mainly discuss how to design a feature extraction network with higher performance and efficiency.
Extracting Optical flow feature maps by using FlowNet 2.0 can effectively filter out background information, to avoid model to learn unnecessary information (including people appearance and background). Furthermore, we added YOLOv2 as the people detector, pruning the excess input size and automating the manual marking process.
In order to overcome the problem that 2D convolutional networks have difficulty in obtaining regional temporal feature, we have proposed a method to connect 3D/2D convolutional structures so that the network has better performance.
