基于多域对抗学习的无人机目标跟踪算法

UAV target tracking algorithm based on multi-domain adversarial learning

针对无人机视频跟踪中正样本不足和单帧强判别特征易导致分类器过拟合的问题,提出一种基于多域对抗学习的实时无人机目标跟踪算法。将生成对抗网络引入到多域学习的特征生成中,利用对抗学习提高特征提取的鲁棒性;在卷积层中加入具有不同扩展系数的空洞卷积进行多尺度特征抽取,构建具有不同感受野的特征提取模块;在交叉熵损失函数中添加调制因子解决正负样本数量不平衡的问题。实验结果表明,该算法的跟踪精度、成功率均得到了提高。

A real-time UAV(unmanned aerial vehicle)target tracking algorithm based on multi-domain adversarial learning was proposed to solve the problems of insufficient positive samples and single frame strong discriminant features which lead to the over fitting of classifier.To improve the robustness of feature extraction,the generative adversarial network was introduced into the feature generation of multi-domain learning.The dilated convolution with different expansion coefficients was added to the convolution layer for multi-scale feature extraction.The feature extraction module with different receptive fields was constructed.The modulation factor was added to the cross entropy loss function to solve the problem of unbalanced number of positive and negative samples.Experimental results show that the tracking accuracy and success rate of the algorithm are improved.