Research On Gait Recognition Based On Deep Learning And Gait Energy Image

Gait recognition aims to identify individual walking patterns in a long-distance,which is one of the most promising video-based biometric technologies.It has enormous potential in crime investigation,access control and social security.However,the variations like different camera viewpoints,complex backgrounds,coat-wearing,which can cause dramatic changes in the appearance.Compared with the face,fingerprint and iris,gait recognition has significant challenges in the real-world.Due to the high effectiveness and reduce tedious computations,Gait energy image(GEI)is utilized in the research field of gait recognition.However,the performance of gait recognition is directly related to gait cycle detection.Most existing methods of gait cycle detection cannot accurately detect the gait cycle in all viewing angles.And a single GEI only focuses on the static characteristics of gait,which ignores the dynamic information.In order to alleviate these issues,we in this project address gait recognition problem based on GEI by employing deep learning.The contributions of this paper can be summarized as follows:Firstly,different from existing methods of gait cycle detection which obtain less precise in cross-view situations,we propose two new gait cycle detection based on classification and regression fitting,respectively.Instead of directly utilize gait geometric feature information,these two methods aim to define the position of the gait as the category value and the sine function value.Then,we can obtain the convolutional neural network model via large training data,which converts the gait period detection into a classification problem or regression simulation.Compared with the existing geometric feature-based methods with the cross-view,the experiments show that our proposed methods obtain better recognition results.However,these methods still cannot achieve high accuracy.Secondly,some researches on the traditional gait period detection and convolutional neural network cannot obtain high accuracy in all viewing angles.Due to the frame rate of different cameras is different in the actual scene,the dynamic poses of the person gait cannot completely capture on the starting and ending positions.This makes it impossible to accurately obtain the gait cycle.Therefore,different from the methods to preprocess gait period detection,we propose a non-periodic gait energy image construction method.Through the analysis of gait movement characteristics and the image experiment pairing with traditional gait energy images,we analyze how to select a fixed number of frames or a fixed duration.The results show that non-periodic gait energy image is expressed one period gait feature information when he fixed number of frames is more than 45.It can achieve the similarly effectiveness to the the traditional methods based on gait energy image.Thirdly,compared with the methods based on singe GEI which only represent one or two gait features,we in this paper propose a multi-template color gait energy image construction framework.The gait energy image such as GEI,AEI and GEnI are fused to build a multitemplate color gait energy.The experimental results show that the multi-template color gait energy image has the ability to capture more gait features.According to the different features in the different parts of the human body,the multi-template color gait energy image is divided into three regions including head,torso and leg regions.These regions are sent into the designed network to extract the gait information,respectively.The final features are fusion to achieve gait recognition.The experimental results show that our method not only retains the overall contour information,but also capture more discrimination local features.Our network has good robustness in some situations such as bag-carrying,coat-wearing and camera viewpoints.Finally,most methods based on based on extracting the spatial features of a single gait image,which are easy ignoring temporal information of the gait sequence.Additionally,the traditional GEI and non-periodic gait energy image need enough frames of a gait sequence,which cannot consider the incomplete cycle of the real scene.Therefore,we design a novel temporal gait energy image method based on deep neural network and attention mechanism.Specifically,the sequential gait energy image is obtained by fixing intervals the multi-template color gait energy image framework,then we extract the high-level features by adopting the modified convolutional neural network.The high-level spatial features of the sequential gait energy image are sent into the channel and spatial hybrid attention network,which enhances local features of gait recognition.The extracted information is sent into LSTM to extract the temporal feature of GEI.We utilize the attention model to enforce temporal features to achieve gait recognition.The experimental results show that our proposed method has the enable to capture the temporal and spatial feature,which can further improve the performance of gait recognition.