Research On Gait Recognition Algorithm Based On Deep Learning

Gait recognition is an emerging technology that uses biological characteristics for recognition.The purpose is to recognize the identity through the different postures of people walking.Compared with biometric technologies such as the face,the fingerprint and the voiceprint,gait recognition has the advantages of non-contact long-distance recognition and not easy to disguise,especially having significant advantages in intelligent video surveillance and other fields.However,in terms of technological development,gait recognition also has limitations such as being vulnerable to environmental influences and weak robustness.In response to this problem,this paper aims to proposing a gait recognition method that combines appearance contours and posture key points to achieve high-precision gait recognition in actual scenes.Firstly,this paper studies the key feature extraction methods in the early stage of gait recognition,focusing on the analysis of three human contour segmentation methods:background subtraction,optical flow and semantic segmentation,and finally determines the semantic segmentation method by comparing the advantages and disadvantages.Through experiments on the public gait data set CASIA-B,it is found that the gait recognition method based on appearance contour is easily affected by the external environment.Therefore,this paper uses the HRNet network to estimate the posture and extract the skeleton key point features and appearance that are not easy to change.Contour features are fused and used together for gait recognition.Secondly,this paper designs a gait recognition network based on fusion features,using convolutional neural network extraction features,compression operation,multi-scale fusion features,horizontal pyramid pooling structure,and fusion of triplet loss function triplet and cross-entropy loss function softmax used for error calculation and training.Experiments show that the algorithm is more advanced than the current gait recognition algorithm in the normal walking state NM,backpack walking state BG and walking state CL in the average rank-1recognition rate can reach 95.5%,89.7%,75.5% respectively,increased by 0.5%,2.5%,5.1%respectively.The model has strong robustness.Finally,the gait recognition algorithm designed in this paper is applied to the campus shooting video,and the gait data set in the real campus scene is collected and produced according to actual needs,the early gait feature extraction network is appropriately changed to take into account the accuracy and speed.It is required that the speed reaches 35 fps and can process pictures in real time.The final experiment shows that the algorithm designed in this paper also has a higher rank-1 recognition rate and strong model applicability on the data set in the campus scene.