全天实时跟踪无人机目标的多正则化相关滤波算法

相关滤波算法(Correlation filter, CF)已广泛应用于无人机目标跟踪. 然而, 受无人机 (Unmanned aerial vehicle, UAV) 平台本身计算性能的制约, 现有的无人机相关滤波跟踪算法大都仅采用手工特征来描述目标的外观, 难以获得目标的全面语义信息. 并且这些跟踪算法仅能较好地进行光照条件良好场景下的跟踪, 而在跟踪夜间场景下的目标时性能严重下降. 此外, 相关滤波跟踪器采用余弦窗口来抑制循环移位产生的边界效应, 缩小了样本提取区域, 产生了训练样本污染的问题, 这不可避免地降低了跟踪器的性能. 针对以上问题, 提出全天实时多正则化相关滤波算法(All-day and real-time multi-regularized correlation filter, AMRCF)跟踪无人机目标. 首先, 引入一个自适应图像增强模块, 在不影响图像各通道颜色比例的前提下, 对获得的图像进行增强, 以提高夜间目标跟踪性能. 其次, 引入一个轻量型的深度网络来提取目标的深度特征, 并与手工特征一起来表示目标的语义信息. 此外, 在算法框架中嵌入高斯形状掩膜, 在抑制边界效应的同时, 有效避免训练样本污染. 最后, 在5个公开的无人机基准数据集上进行充分的实验. 实验结果表明, 所提出的算法与多个先进的相关滤波跟踪器相比, 取得了有竞争力的结果, 且算法的实时速度约为25 fps, 能够胜任无人机的目标跟踪任务.

All-day and Real-time Multi-regularized Correlation Filter for UAV Object Tracking

Correlation filter (CF) has been widely used in unmanned aerial vehicle (UAV) object tracking. Due to the computational limitation of the UAV platform, the existing UAV tracking algorithms rely heavily on the hand-crafted features, which cannot obtain all the semantic information of the target. Meanwhile, the existing tracking algorithms focus on the tracking of daytime targets, and ignore the tracking problem of nighttime targets. In addition, when the correlation filter-based tracker uses the cosine window to suppress the boundary effect caused by the cyclic shift, the sampling area is shrunk, which causes the contamination of training samples and inevitably deteriorates the performance of the tracker. Aiming at the above problems, we propose an all-day and real-time multi-regularized correlation filter (AMRCF) for UAV object tracking. Firstly, an adaptive image enhancement module is introduced to enhance the image without affecting its color ratio of each channel and improve the nighttime tracking performance. Secondly, a lightweight deep network is introduced to extract the deep features of the target and represent the semantic information together with the hand-crafted features. Thirdly, a Gaussian-shaped mask is embedded in the correlation filter framework, which can effectively avoid the contamination of training samples while suppressing the boundary effect. Finally, extensive experiments are conducted on five publicly available UAV benchmark datasets. The experimental results show that the proposed algorithm achieves competitive result compared with several state-of-the-art correlation filter-based trackers, and the real-time speed is about 25 fps, which makes it competent for UAV object tracking task.