基于特征融合与通道感知的无人机红外目标跟踪算法

针对手工提取特征对红外目标不敏感,导致无法准确跟踪红外目标的问题,在全卷积孪生网络框架下,融合多层深度特征并结合通道感知提出了一种无人机红外目标跟踪算法。首先使用预训练网络提取目标深度特征,分别提取待跟踪目标的Conv4-1、Conv4-3、Conv5-2层特征,进而通过梯度计算选择对于目标活动和尺度变化较为敏感的特征通道参与后序的互相关操作,并通过计算模板图像和候选区域搜索图像之间的相似度获取目标响应图。最后利用平均峰值相关能量对跟踪结果进行评估并使用卡尔曼滤波对跟踪结果进行修正。在LSOTB-TIR红外目标跟踪数据集上进行了性能测试并与当前九种优秀的算法进行了对比,实验结果表明,本文算法跟踪成功率最高,能够有效应对红外目标跟踪中热交叉、干扰源等挑战,且具有较好的实时性。

UAV infrared target tracking algorithm based onfeature fusion and channel awareness

To address the problem that manually extracted features are not sensitive to infrared targets,resulting in inaccurate tracking of infrared targets,a UAV infrared target tracking algorithm is proposed in the framework of fully convolutional Siamese networks,fusing multi layer depth features and combining with channel perception.Firstly,the pre training network is used to extract the target depth feature,and the Conv4 1,Conv4 3 and Conv5 2 layers of the target to be tracked are extracted respectively.Then,the feature channels that are more sensitive to the target activity and scale changes are selected through gradient calculation to participate in the subsequent cross correlation operation,and the target response map is obtained by calculating the similarity between the template image and the candidate region search image.Finally,the average peak correlation energy(APCE)is used to evaluate the tracking results,and the Kalman filter is used to correct the tracking results.The performance test is carried out on the LSOTB TIR infrared target tracking data set,and compared with the current nine excellent algorithms.The experimental results show that the algorithm in this paper has the highest tracking success rate,can effectively deal with the challenges of thermal crossover and distractor in infrared target tracking,and has better real time performance.