基于转换残差的行人重识别研究

Research on Person Re-Identification Based on Transformed Residuals

行人重识别可以跨摄像头识别行人身份,在城市治安管理和智能交通系统等领域发挥着重要作用。然而,行人重识别任务存在着风格差异,行人全局和局部特征之间的关系有待进一步挖掘。为此,提出基于转换残差的行人重识别算法。转换残差模块通过实例归一化改善风格差异,同时通过卷积操作捕获更丰富的图像细节特征。该算法引入特征融合模块,通过将Transformer块和转换残差模块残差连接,综合利用两者特征,实现全局和局部特征的有效融合。基于该算法在Market1501和MSMT17数据集上进行实验,平均精度(mean Average Precision,m AP)评价指标分别达到了92.1%和69.6%。

Person re-identification enables the identification of individuals across different cameras,making significant contributions to city security management and intelligent transportation systems.However,challenges such as style variance and the relationship between a person's global and local features present obstacles in this field.To address these issues,a person re-identification algorithm based on transformed residuals is proposed.The transformed residual module employs instance normalization to mitigate style variance and utilizes convolutional operations to capture richer image detail features.The proposed algorithm features a fusion module that integrates the Transformer block and the transformed residual module through residual connections,effectively combining global and local features.Experiments on Market1501 and MSMT17 datasets demonstrate the algorithm's efficacy,yielding mean Average Precision(mAP)evaluation metrics of 92.1%and 69.6%,respectively.