| With the gradual development of the new era and the advancement of technology,the security of personal identification information has become more and more important.Recognition technologies with biometric features such as face recognition,fingerprint recognition and iris recognition are constantly receiving attention.Although these identification technologies have great advantages over traditional identification technologies,they have many drawbacks,such as face recognition being easily affected by occlusions,angles and other factors,and fingerprints being vulnerable to breakage.Therefore,gait,as a biometric feature of behavior,has become a hot topic of research by experts and scholars in recent years because of its advantages such as being difficult to disguise and no need for human cooperation.However,many problems are still waiting to be solved in the research of gait recognition technology.As the amount of gait frame image data increases,in order to achieve a high recognition rate,it makes the network model achieve gait recognition larger and larger,which makes it difficult to deploy to mobile due to the limited computing and storage capacity of mobile;in gait recognition,the recognition accuracy decreases due to the obscuration of pedestrian clothes and carrying objects.To address the appealing problem,this thesis combines the idea of deep learning and the advantages of deep learning in image classification recognition to carry out gait recognition research,mainly accomplishing the following work.(1)The common networks for gait recognition are analyzed,and it is found that the traditional convolutional neural network has many layers and the number of training models is large.In view of the above problems,the traditional lightweight network Mobile Net V1 is optimized and improved,and the lightweight network Mobile Net V1 model incorporating the attention mechanism is proposed.By using the gait energy image as the input of the network and incorporating the attention CBAM module into the Mobile Net V1 network model with deep separable convolution,the method takes advantage of the simplicity of the Mobile Net base network module and then incorporates the attention mechanism to enable the network to autonomously learn the important regions of the image and suppress the minor regions,which improves the recognition rate and reduces the computational power.Thus the network recognition performance is improved.(2)To address the problem of occlusions such as pedestrian clothing and carrying objects,which affect the accuracy of gait recognition,it was found that gait energy images are commonly used as feature representations for gait recognition.Through the analysis of the gait energy image synthesis method,it is found that the gait energy image ignores the interrelationship between frames when synthesizing,resulting in the loss of timing information.To address this problem,this thesis proposes an improved gait energy image,which maps the temporal information on the human skeleton by means of RGB color mapping to attenuate the effects of covariates such as changes in viewpoint and carry-over occlusion for gait recognition,and then fuses the contours with the skeleton containing temporal information to obtain an improved gait energy image,which is also used as the input of the network and fed into the fused attention mechanism in the lightweight network to perform gait recognition.(3)The proposed method is evaluated in the CASIA-B dataset,which is publicly available at the Institute of Automation,Chinese Academy of Sciences,for experiments.The experimental results show that the proposed method has a good recognition effect under various covariates,and achieves good recognition accuracy and significant improvement in recognition rate by comparing with other methods,which verifies the effectiveness of the proposed algorithm. |
