【摘要】 针对基于深度学习的车辆重识别模型缺乏提取车辆局部细节描述的能力,以及不同环境下同一车辆表观特征变化显著,增加车辆重识别的难度的问题,提出一种融合多尺度车辆表观和属性信息的重识别模型.充分利用车辆颜色和车型属性在不同环境下较为稳定且易判断的特性,将其融合到车辆表观特征中,获得强化提升的深度特征;另外使用交叉熵损失函数与Triplet损失函数构建多任务的目标优化函数对模型进行参数训练.该方法在VeRi-776车辆数据库上的实验结果表明:通过融合颜色和车型属性特征可以明显地提高车辆重识别的准确率,并可以取得优于其他大部分对比方法的性能.更多还原

【Abstract】 Although global features extracted from the deep learning-based model have secured strong discrimination abilities,they endure lack of descriptions about local details.Besides,the appearance of the same car in different environments changes significantly,thus increasing the difficulty of vehicle re-identification.Therefore,in this study,we propose a vehicle re-identification model which integrates multi-scale vehicle appearance and attribute information.Due to the robustness of the attribute information in different environments,we fuse attribute features into the appearance features to obtain an enhanced feature for accurate vehicle reidentification.Finally,a multi-task loss function containing the cross-entropy loss and the triplet loss is adopted to train model parameters.We conduct extensive experiment on the large-scale VeRi-776 dataset,and experimental results demonstrate the effectiveness of our proposed method,which can significantly improve the accuracy and outperform most compared methods.更多还原