A Study Of Gait Recognition Based On Human Pose Feature

The gait feature is a kind of biometrics,which has the characteristics of distance acquisition,non-contact and not easy to camouflage.Especially under surveillance,the distance between the camera and the pedestrian is far.The common feature like face,fingerprint and iris are not available,and gait feature which is pedestrian walking attitude is a more feasible method.Many researchers have demonstrated that gait recognition is a very effective method of identification at long distances.Therefore gait recognition is worth to research and has a great potential in the real application.There have been mainly two different categories of gait methods along with the steady progress of gait recognition.The appearance-based is one of the most popular methods in the gait recognition task,and some methods have reached relatively high recognition rates with cross-view variation,cross-clothing variation and so on.Nevertheless,human silhouettes are sensitive to view angles,clothing variations and the carrying conditions.The second category of gait recognition method is the model-based method.The model-based methods model human body structure and local movement patterns of different body parts.It is generally robustness to the cross view and shape changes.However,most model-based methods generally need to manually mark the data or use specific device to obtain the human joint information.It is really a challenging task to localize human body joints accurately.In order to solve these two challenges,we propose a novel pose-based gait recognition method which is robust in the clothing and carrying conditions.In addition,we propose a pose-based temporal-spatial network(PTSN)to extract the temporal-spatial features,which effectively improve the performance of gait recognition.The neural network used in the proposed method combines Long Short Term Memory(LSTM)recurrent neural network with Convolutional Neural Network(CNN).LSTM can learn personal identity from the temporal information,and CNN can learn from human body static information.In order to extract the temporal-spatial features with large inter-class variations and reduce the intra-class variations,we adopt a multiloss strategy to optimize the PTSN network.Experiments evaluated on the challenging CASIA B show that the pose information can help to improve gait recognition greatly in terms of bag and clothing conditions.But it is still challenging in the crow-view condition.In order to solve the problem of cross-view,we try to discover and verify the capability of the 3D pose on view invariant gait recognition,and show the unique advantages of 3D model based methods.In these several years,the progress on deep learning makes human pose estimation feasible,and the 3D pose can be estimated accurately.The 3D human pose naturally has greater capability on view invariant than the 2D pose and appearance based features.We evaluated on the challenging dataset CASIA B and achieve comparable results with the popular cross-view methods under a uniform model which could handle cross-view conditions.The results show the proposed has advantages in large view variance and it has great potential to fuse with the appearance based methods.