Research On Deep Learning Gait Recognition Method Integrating Spatiotemporal Information

The Pingan City and Xueliang Project promoted by the Chinese government can provide reliable and efficient technical support for public safety construction.Among them,gait recognition technology is a technology that can assist supervisors to track or hunt down suspect.However,existing gait recognition methods are susceptible to changes in external environment and walking conditions,posing challenges in practical applications.For this reason,with the support of the National Natural Science Foundation of China,this thesis has carried out research on deep learning gait recognition method integrating spatiotemporal information,aiming to improve the accuracy of gait recognition under different perspectives,different clothing conditions,and infrared states.The innovative research achievements and main work completed in this thesis are as follows:(1)Multi-view gait cycle detection by fitting geometric features of lower limbs.Gait cycle detection is the core part of gait recognition tasks,and its detection results directly affect the accuracy of gait recognition.However,the existing gait cycle detection methods have the limitation that the detection effect depends on the shooting angle.Therefore,a multi-view gait cycle detection method fitting the geometric features of lower limbs is proposed.Firstly,use the BlazePose pose estimation algorithm to extract the human pose topology map to simplify the image preprocessing process.Then,by analyzing the periodic dynamic change law between the joint points in the human posture topology map under walking state,the inclination formed by the left shin and the horizontal ground and the Euclidean distance ratio from the midpoint of the left and right hip joints to the left and right ankle are extracted as features.Finally,the feature data were fitted into sinusoidal function waves by Fourier transform,and the gait period is detected based on the fitting results.On the CASIA-B dataset,the ratio value of the frame error of the proposed method to the actual gait cycle is 0.13,indicating that the proposed method has achieved good front and back view and strabismus angle detection results.(2)Deep learning gait recognition based on two branch spatiotemporal feature fusion.Aiming at the problem that the existing gait recognition methods are easily affected by shooting angle and clothing changes,this thesis proposes a deep learning gait recognition method that fuses 2D shoulderless pose topological energy maps(SPTEM)and 3D local skeleton gait features(LSGF).Firstly,the lightweight BlazePose pose estimation algorithm is used to extract the human posture topology to generate the SPTEM,which improves the detection speed and reduces the impact of clothing changes.Then,LSGF is introduced to make up for the low recognition accuracy deficiency of a single energy map feature in the case of variable viewing angles.Finally,a spatio-temporal feature extraction network model fused with an attention mechanism is proposed,and the dual-stream features are fused uniformly in the fully connected layer.The proposed algorithm is validated on the CASIA-B dataset and compared with the current mainstream gait recognition methods.The results show that the gait recognition rate of this method is significantly improved under cross-view and cl conditions.(3)Dual-stream deep network infrared gait recognition based on residual multi-scale fusion.A residual multi-scale dual-stream network model based on silhouette differential fusion flow and silhouette flow is proposed to address the issue of convolutional neural networks not being able to fully capture and utilize spatiotemporal information in the gait recognition of low-quality infrared images.Firstly,a fine-grained segmentation strategy combining Faster-RCNN and Deeplab v3+ algorithms is used at the input end of the model to extract silhouettes,reducing the impact of noise information and avoiding feature loss.Secondly,add a gait silhouette differential fusion module to the branch network on the model to obtain the difference and change information between adjacent silhouette frames.Then,residual units and multi-scale feature fusion techniques are used in the feature extraction section of the model to deepen the network hierarchy and extract spatiotemporal information of different granularity,respectively.Finally,the multi-scale pyramid mapping module is used to further enhance the model’s ability to represent local and global features.On the CASIA-C dataset,the average gait recognition rate of the proposed method is 98.85%,which can meet the recognition requirements in infrared scenes.