Research On Gait Recognition Based On Deep Learning

Gait recognition is a biological feature recognition method based on human's gait.It has unique advantages and is widely applied around the world compared to other recognition methods.However,in the research process of gait recognition,there are some problems to be solved,such as gait silhouette variation caused by changing or carrying view angle,missing joint point information in gait template and insufficient utilizing temporal information on gait silhouettes.These problems largely limit the performance of gait recognition.Considering the above mentioned problems,this paper combines the advantages of deep learning classification and recognition,gait recognition based on deep learning to finish the following works:(1)Gait recognition based on improving chrono-gait image and multi-feature triplet network.Dynamic information is missing in the chrono-gait image,and the temporal information is susceptible to view.This paper presents an improved chrono-gait image based on gait silhouette joint positioning,thus an improved chrono-gait image add dynamic information,meanwhile temporal information is mapped on the lower limb skeleton,and the skeleton composed of joints.To reduce the effects of changes in view carriers,the improved gait template will contain more gait information.At the same time,in order to improve the effect of feature similarity learning of triplet network,gait recognition method based on improved chrono-gait image and multi-feature Triplet network is proposed.Sufficiently the network extract the global and local features in improved chrono-gait image.And input fuses features into the metric network for similarity learning,to improve gait recognition accuracy.The experimental results show that improved chrono-gait image has more advantage in gait recognition,Compared with similar methods in recent years,our method raises the whole average accuracy 4.9%-8.2%.It can better overcome the influence of covariates such as carriers and wearing,and improve the accuracy of gait recognition.(2)Gait recognition based multichannel spatiotemporal network and joint optimization lossA gait recognition method based on a gait skeleton sequence is more robust in changing perspectives and carriers,and it also loses a large number of highly discriminative contour features.Its advantages and disadvantages are just complementary to the gait recognition method based on the gait contour sequence,therefore,this paper presents a multichannel spatiotemporal network,gait skeletons and silhouettes sequences as input for multichannel spatiotemporal network.Combined with the attention to sufficiently extract the spatiotemporal features in gait sequences,A gait recognition model is established based on the above metric learning and classification network.This paper further optimizes the selection of positive and negative samples of the triplet loss during the training process,at the same time,the label lmoothing regularization(LSR)integrated to optimize the cross-entropy loss.This paper combines the above two optimization strategies to jointly supervise network training,enhanced metric learning network generalization,improve the accuracy of gait recognition in scenes such as views and carrying.The experimental results show that the proposed method has an average accuracy improvement of 11.5%-14.2% compared with similar methods in recent years,and it has more advantages in view change and carriers.(3)Design gait recognition system based on deep learning.In this paper,the gait recognition based multichannel spatiotemporal network and optimization triplet loss has more robust and higher recognition,so we combine Pytorch deep learning framework and OpenCV library,then a gait recognition system is designed and based on deep learning.Finally we collect gait videos in real life,verify the effectiveness of the gait recognition system and recognition algorithm.