增强特征信息的孪生网络无人机目标跟踪方法

Twin network UAV target tracking method based on enhanced feature information

为解决无人机视觉下目标因光照变化、完全遮挡、快速运动等情况导致跟踪效果变差甚至跟踪失败的问题,基于全卷积孪生网络跟踪算法SiamFC提出一种增强特征信息的目标跟踪方法。采用GOT-10K数据集替换原训练数据集ILSVRC2015-VID对模型进行训练,构造理解能力更深的网络模型;将带有高语义信息和低细节的浅层特征融入到深卷积层中增强网络对目标特征的提取能力;引入轻量级条带池化模块加强目标特征信息。在UAV123公开数据集基准上进行测试,实验结果表明,该方法的成功率和精确度分别达到0.542和0.746。

To solve the problem of poor tracking effect or even tracking failure caused by the change of illumination,complete occlusion and rapid movement of the target in UAV vision,a target tracking method based on full convolution twin network tracking algorithm SiamFC was proposed to enhance the feature information.The original training data set ILSVRC2015-VID was replaced by GOT-10K data set to train the model and construct a network model with deeper understanding ability.The shallow features with high semantic information and low details were integrated into the deep convolution layer to enhance the ability of the network to extract target features.A lightweight stripe pooling module was introduced to enhance the target feature information.Results of experiments on UAV123 open dataset benchmark show that the success rate and accuracy of the method are 0.542 and 0.746 respectively.