针对无人机视频的带干扰抑制的l_(1)正则相关滤波跟踪算法

l_(1) Regularized Correlation Filter Tracking Algorithm with Distractor Suppression for UAV

近来,相关滤波因其精度和鲁棒性良好而在目标跟踪领域获得广泛应用.然而,相关滤波的隐式循环样本带来了严重影响跟踪性能的边缘效应.在之前的工作中,采用l_(1)正则消除了传统相关滤波模型边缘效应的影响,但仍然通过将隐式循环样本回归到一个固定的目标标签来学习一个能区分前景和背景的回归模型.这种预定义的不变的回归标签使得跟踪模型对于不确定性高的航拍跟踪场景的鲁棒性和适应性较低.因而,提出一种利用在检测阶段生成的响应图的局部最大值来自动定位当前帧的干扰物的方法,通过抑制干扰物在回归学习中的响应,可以动态地、自适应地改变回归目标,从而提高跟踪的鲁棒性和适应性.在对比实验中,该方法在DTB70数据集上的精确率和成功率分别达到66.9%和43.4%,相较于基准算法,精确率和成功率分别提高了4.3个百分点和2.1个百分点.在UAV123@10 fps上,精确率和成功率分别达到58.2%和40.9%.实验结果表明,提出的跟踪器的性能优于其他几种代表性的方法.

Recently,correlation filter has yielded the wide applications in visual tracking field because of its good accuracy and robustness.However,the implicit circulated samples of correlation filter introduce the bound⁃ary effects,which severely degrades the performance of tracking model.In previous work,the l_(1) regularization is utilized to alleviate the boundary effects of traditionary correlation filter tracker,which learns a regressor to dis⁃tinguish the foreground from the background by regressing the implicit circulated samples into a fixed target la⁃bel.However,this predefined and unchanged regression target makes the tracking model less robustness and adaptivity to uncertain aerial tracking scenarios.Thus,in this work,local maximum points of the response map generated in the detection phase is exploited to automatically locate current distractors.By repressing the re⁃sponse of distractors in the regressor learning,the regression target can be altered dynamically and adaptively to leverage the tracking robustness as well as adaptivity.In the contrast experiments,the proposed algorithm a⁃chieves the accuracy and success rate at 66.9%and 43.4%on the DTB70 datasets.The success rate and accu⁃racy rate are rose by 4.3 and 2.1 percent points in comparison with benchmark algorithm.And the accuracy and success rate reached 58.2%and 40.9%on UAV123@10fps datasets.The extensive experimental results dem⁃onstrate that the proposed tracking model is superior to other several representative methods.