基于优化GaitSet模型的步态识别研究

Research on gait recognition based on optimized GaitSet model

为了能够保留更多的输入步态轮廓图的轮廓和纹理等信息,增加深浅特征融合的输入,提出了基于优化GaitSet模型的步态识别研究.首先,在原模型的基础上,采用了联合Loss优化,步态轮廓图的裁剪对齐为较大图片尺寸,并且进行进一步的论证;其次,增加多尺度的感受野的特征输入和深浅特征融合;最后,在CASIA-B数据库上进行了验证.实验结果表明:LT样本规模及非相同视角下,NM状态下的识别率为97.309%,BG状态下的识别率为94.048%,CL状态下的识别率为81.736%.相比较原模型Rank-1的准确率有较大的提升.

In order to retain more information such as the contour and texture of the input gait contour map and increase the input of deep and shallow feature fusion,a research on gait recognition based on the optimized GaitSet model is proposed.The optimization content mainly includes the following parts:First,on the basis of the original model,the joint Loss is used to achieve better results.Then,the crop alignment of the gait contour map is that the larger image size contains more texture and contour information,which has a greater impact on the recognition rate,and further demonstration is carried out.Secondly,increase the feature input of multi-scale receptive fields and the fusion of deep and shallow features.Combined with the HRNet module,it has strong robustness to increase the input of multi-scale features while maintaining a higher resolution,some other adjustments,verify on the CASIA-B database.The experimental results show that under the LT sample size and different viewing angles,the recognition rate in the NM state is 97.309%,the recognition rate in the BG state is 94.048%,and the recognition rate in the CL state is 81.736%.Compared with the original model Rank-1,the accuracy rate is greatly improved.