结合交并比损失的孪生网络目标跟踪算法研究

SiamRPN算法采用Ln范数损失训练边界框预测，未考虑预测框与真值框间交并比(inersection over union, IoU)的关系，导致准确性不足。针对该问题，提出一种结合IoU损失的SiamRPN目标跟踪改进算法。设计了IoU-smooth L1范数联合优化模块，对候选正样本进行IoU损失与smooth L₁范数损失的联合优化；依据回归预测结果，用预测框与真值框的IoU作为权重对正样本进行加权分类预测，增加正样本间的区分度，同时确保分类预测与回归预测的关联性。对比实验结果表明：本文所提改进算法能有效提升跟踪性能。

Siamese Object Tracking Algorithm Combined with the Intersection over Union Loss

To improve the accuracy of the object bounding box regression prediction of the SiamRPN,solve the problem of low discrimination of positive samples in classification prediction and the lack of correlation between regression prediction and classification prediction, an improved object tracking algorithm of SiamRPN which combined with IoU(intersection over union) loss is proposed. A joint optimization module of IoU-smooth L1 is designed to optimize the IoU loss of the best positive sample and the smooth L1 loss of other positive samples jointly. According to the regression prediction results, the weighted classification prediction is performed on the positive samples with the weight calculated by the IoU of the prediction box and the truth box, so as to increase the discrimination between the positive samples, while ensuring the correlation between the classification prediction and the regression prediction.The results show that the proposed algorithm can effectively improve the tracking performance.