基于自适应深度网络的无人机目标跟踪算法

针对无人机（UAV）视频中目标易受到遮挡、形变、复杂背景干扰等问题，提出一种基于自适应深度网络的无人机目标跟踪算法。首先，基于主成分分析（PCA）和卷积神经网络（CNN）算法，设计3阶的自适应深度网络进行目标特征提取，该网络对图像的H、S、I通道分别进行主成分分析学习，将得到的特征向量输入网络进行分层卷积，优化了网络结构，提高了网络的收敛速度和精度。其次，将目标深度特征输入核相关滤波算法进行目标跟踪，通过分析相邻2帧图像的变化率，采用分段自适应调整学习率的算法进行目标模板更新，有效地改善目标遮挡问题。仿真实验结果表明，该算法有效地避免了复杂因素干扰导致的跟踪精度下降，具有较好的鲁棒性，相较于全卷积跟踪（FCNT）算法平均跟踪精度提高了9.62%，平均跟踪成功率提高了11.9%。

UAV target tracking algorithm based on adaptive depth network

Aiming at the problem that targets are subject to occlusion, deformation, and complex background interference in the drone video, a Unmanned Aerial Vehicle (UAV) target tracking algorithm based on the adaptive depth network is proposed. First, based on the Principal Component Analysis (PCA) and Convolutional Neural Network (CNN), a 3-order adaptive CNN network is designed for target feature extraction. PCA is hierarchically performed on H,S, and I channels, convolving hierarchically by the obtained eigenvectors, which optimizes the network structure and improves the convergence speed and accuracy. Second, the target depth feature is input into KCF algorithm for target tracking. By analyzing the change rate of the two adjacent frames and using the segmented adaptive adjustment of learning rate to update the target template, the target occlusion problem is effectively moderated. The experimental results show that the algorithm effectively avoids the degradation of tracking accuracy caused by complex factors, reaching good robustness. The average accuracy-rate of the algorithm is 9.62% higher than that of fully convolutional network based tracker Fully Convolutional Network Tracking (FCNT), and the average success-rate is increased by 11.9%.