Interpretable Gait Recognition Based On Optimal Dynamic Geometric Features

Gait recognition is a new biometric recognition technology.Compared with the existing recognition technologies such as face,fingerprint and iris,gait recognition has the advantages of long-distance,non-contact,non-cooperation and concealment.In recent years,it has attracted extensive attention from researchers in the field of computer vision and biometric recognition.These advantages of gait recognition are urgently needed in security prevention and control,suspect tracking and smart city.Although gait recognition technology has developed rapidly in recent years,there are still some difficulties to be overcome if it wants to be applied in reality,such as the change of viewing angle,the change of pedestrian wearing state,the change of carrying objects and so on.In addition to the above poor robustness to covariate factors,the user’s confidence in the gait recognition results is also a difficulty to be considered in the landing use of gait recognition.Therefore,there is still a way to go to apply gait recognition in reality like other biometric recognition technologies.In this paper,some difficulties of gait recognition are studied as follows.In view of the loss of human structure information and poor interpretability caused by not paying attention to the unique motion patterns of different human parts in the use of binary contour map of gait sequence,this paper uses the method of example segmentation to distinguish different parts of human body,so as to obtain the segmentation contour map of human parts,While increasing the performance of gait recognition,it also increases the interpretability of gait recognition task.Three different human body part differentiation methods are proposed to explore different application methods of human body part segmentation contour map.For the temporal and spatial information in gait contour sequence,different depth and multi-stage feature fusion methods are used to extract,and attention mechanism is used in different stages of fusion to ensure richer feature fusion.After that,the fused gait spatio-temporal features are expressed by pyramid mapping.The experimental results show that this human part segmentation contour map performs very well in gait recognition task.In view of the lack of interpretability of the existing gait recognition research,this paper proposes an end-to-end optimal dynamic geometric feature model,which can well explain which local areas of the human body are the most critical in the task of gait recognition while ensuring the accuracy of gait recognition.Firstly,the obtained human key points are expanded according to the rigid structure and body structure of human bones,and the position of human key points is dynamically optimized by 3D convolution neural network.Then,the gait contour is divided into local areas of human body with time invariance by using the expanded and optimized human key points,The index of important local areas of human body is inferred and selected through the key area inference model,so as to obtain the local binary contour map of gait sequence,send the inferred local binary contour map into the gait recognition model for recognition,and use the end-to-end joint training of key area inference model and gait recognition model to update the model parameters iteratively,By gradually reducing the selected human parts,we can gradually get the most favorable human parts for gait recognition,so as to better explain gait recognition.