Research On Key Techniques Of Gait Recognition Based On Spatio-Temporal Multi-Scale Convolution

Gait recognition is a kind of biometric recognition technology that identifies pedestrians according to their posture characteristics.Due to its non-contact characteristic,it can be recognized at a long distance,which has a wide range of application requirements in criminal investigation and forensic identification.However,the application scenarios of gait recognition technology are complex and easily affected by factors such as clothing,perspective and carried objects.It is a challenging task to improve the recognition rate.In order to improve the performance of gait recognition algorithm in various scenarios,the main work of this paper is as follows:(1)Proposed a gait recognition method based on spatio-temporal multi-scale convolution and frame attention mechanism.Aiming at the problem that the existing methods are not enough to extract spatio-temporal features,a multi-temporal scale feature aggregation module is designed to fuse sequence features from multiple different temporal scales to adapt to the complex diversity of human motion patterns.In space,a multi-spatial scale feature extraction module is designed.A parallel multi-resolution convolutional network is used to extract gait spatial features at different spatial scales,and a meter-based frame attention module is proposed to analyze the quality of each frame and generate corresponding weights to highlight the key frame information in the sequence.Experiments were carried out on two popular datasets.The recognition accuracy of the model in CASIA-B dataset under three conditions of normal walking,backpack walking and coat walking are 97.3%,93.4%and 83.1%respectively,and the recognition accuracy on OU-MVLP dataset is 89.9%,which shows the effectiveness and universality of the algorithm.(2)Proposed a gait recognition method based on multi-modal learning and multi-scale time adaptive convolution.Aiming at the problem of low accuracy of existing gait recognition models in complex scenes,a multi-modal feature extraction module is designed,which takes the data of two modalities as input to extract rich spatio-temporal features from the appearance contour of pedestrians,and uses the ability of posture features to resist occlusion to extract more discriminative gait features.Aiming at the problem that the existing models use fixed temporal convolution kernels,which cannot extract temporal features flexibly,a time adaptive modeling strategy is introduced to generate the corresponding temporal convolution kernels for each specific gait video sequence,and different expansion coefficients are given to the temporal convolution kernels in the subsequent convolution operations,to analyze the motion patterns of various parts of the human body more flexibly.Experiments on the CASIA-B dataset show that the recognition rates of the model under three walking conditions are 97.5%,93.6%and 83.9% respectively,which shows the effectiveness of the method in complex scenes.