【摘要】 为了解决核相关滤波跟踪算法在复杂场景下跟踪效果差的问题,提出了一种融合深度特征和尺度自适应的相关滤波目标跟踪算法。首先,通过深度残差网络(ResNet)提取图像中被跟踪区域的深度特征,再提取目标区域方向梯度直方图(FHOG)特征,通过核相关滤波器学习,分别得到多个响应图,并对响应图进行加权融合,得到跟踪目标位置。其次,通过方向梯度直方图(FHOG)特征,训练一个PCA降维的尺度滤波器,实现对目标尺度的估计,使算法对目标尺度发生变化有很好的自适应能力。最后,根据响应图的峰值波动情况改进模型更新策略,引入重新检测机制,降低模型发生漂移概率,提高算法抗遮挡能力,在标准数据集OTB100中与其他7种目标跟踪算法进行比较。结果表明,相比原始KCF算法,改进后的KCF算法精准度提升了29.4%,成功率提升了25.9%。所提算法实现了对跟踪目标位置的精准估计,提高了尺度自适应能力和算法速度,增强了算法抗遮挡能力。更多还原

【Abstract】 Aiming at the problem of poor tracking by the kernel-related tracking filter algorithm in complex scenes, proposed a correlation filter target tracking algorithm combining depth features and scale adaptation.Firstly, the deep residual network(ResNet) was used to extract the depth features of the tracked area in the image, and then the target area directional gradient histogram feature(FHOG) was extracted, and multiple response maps were obtained through the kernel correlation filter learning, and were performed weighted fusion to obtain the tracking target position.Secondly, a PCA dimensionality reduction scale filter was trained through the directional gradient histogram(FHOG) feature to realize the estimation of the target scale, so that the algorithm had a good adaptive ability to the change of the target scale.Finally, according to the peak fluctuation of the response graph, the model update strategy was improved and the re-detection mechanism was introduced to reduce the probability of model drift and improve the anti-occlusion ability of the algorithm.Compare with other 7 target tracking algorithms in the standard data set OTB100.The experimental results show that the accuracy of the original KCF algorithm is improved by 29.3%,and the success rate is improved by 25.3%.The proposed algorithm achieves accurate estimation of tracking target position, improves the scale adaptive ability and the speed of the algorithm and enhances the anti-occlusion ability of the algorithm.更多还原