Gait Identity Recognition Based On Improved Residual Networ

Gait recognition was one of the most promising recognition technologies in the current biometric field.Compared with other biometric features such as face and fingerprint,gait had the advantages of not being limited by distance and not requiring proximity to the target,which maked gait-based identification gain wide attention;gait recognition technology had also flourished in the security field,criminal investigation and other fields.In practical scenarios,due to the high number of layers of mainstream recognition algorithms,the single video input situation leaded to the difficulty of deploying recognition systems on light surveillance devices.At the same time,existing gait identification methods might be affected by unfavorable factors such as wearing and accessory items with viewpoint.In order to solve the above problems,this paper focused on gait recognition based on residual networks,and the specific work was as followed:Res Gait,a gait identity recognition algorithm based on residual networks,was proposed to improve the recognition accuracy of the model.Firstly,the gait dataset was preprocessed and the format was unified using Takemura cropping method.Secondly,the residual module M-Resnet was proposed to modify the channel width and activation function of the residual network,and the size of the convolutional kernel in the first layer of the two-fold wide two-channel residual network was adjusted and optimized to remove a large number of redundant features and finally achieve the minimum overall error.Then,the sampled walking video sequences of subjects were used for feature extraction and fused with the pedestrian image sequences with spatio-temporal characteristics,and the improved residual module was used as the input of the Res Gait network to extract spatial features by parallel spatial convolution and output in the form of 3D vectors,which solved the problem that the spatio-temporal continuity of gait was difficult to be preserved.Finally,after gradient descent and back propagation,the Res Gait network model was optimized by combining batch all and triplet loss functions to improve the accuracy and computational speed of the algorithm.The results showed that Res Gait achieves 96.7%,93.6% and 83.6% recognition accuracy for three walking states on CASIA-B dataset: walking normally,carrying backpack accessories and wearing a coat,respectively,which could prove the practicality and efficiency of the network model with the addition of the improved residual network.On OU-MVLP,the recognition accuracy under multiple views also reached 89.7%,while the results of Res Gait direct tested on homemade datasets were all higher than the current mainstream algorithms.The experimental data demonstrated that gait identity recognition of pedestrians based on improved residual networks was feasible.The multi-person gait recognition algorithm combining YOLOv5 and Deepsort detection tracking was proposed to improve the gait recognition accuracy in multi-person environment and solved the problem of mismatch between the detection module and the output identity of gait recognition.Firstly,the feasibility of the multi-person gait identity recognition algorithm was verified by comparing classical recognition algorithms using gait similarity as well as recognition rate.Secondly,the similarity of each subject’s feature vector dictionary was distinguished and the identity ID was recognized using the hypothesis testing principle as the judging criterion,and finally tested on a large gait dataset OU-MVLP,and an average recognition accuracy of86.9% was obtained.It could be proved that the algorithm had certain real-time and practicality in the face of complex environment with medium and long distance.