基于空间置信蒙版与PCA-HOG特征的目标跟踪算法
|
摘要
为解决核化滤波器存在目标形状限制以及循环矩阵造成训练结果不真实的问题,文中的目标跟踪算法引入CSR-DCF(Discriminative Correlation Filter Tracker with Channel and Spatial Reliabilit)算法中的空间置信蒙版作为位移相关性滤波器,来适应不规则形状的目标。引入PCAHOG(Principal Component Analysis of Histogram of Oriented Gradient)特征来优化尺度相关性滤波器,在不造成HOG(Histogram of Oriented Gradient)特征信息损失的同时降低特征数据的维度,并利用投影矩阵加速傅里叶变换的计算。实验使用VOT2018的视频集以及评估标准,结果表明,改进后的算法在鲁棒性上优于CSR-DCF。
Considering target shape restriction and the training result distortion of the Kernelized Correlation Filters,the object tracking algorithm in this paper introduces concepts of channel and spatial reliability of CSR-DCF( Discriminative Correlation Filter Tracker with Channel and Spatial Reliability) as the translate correlation filter,in order to improve the tracking of non-rectangular objects. The PCA-HOG( Principal Component Analysis of Histogram of Oriented Gradient) features are also used to improve the scale correlation filter by reducing the dimensions of HOG( Histogram of Oriented Gradient) features without information loss,and accelerate the calculation of fourier transform by using projection matrix.The evaluation of the experiments is based on VOT2018 benchmarks and video sequences. The results show that the improved algorithm performs better robustness than CSR-DCF.
Considering target shape restriction and the training result distortion of the Kernelized Correlation Filters,the object tracking algorithm in this paper introduces concepts of channel and spatial reliability of CSR-DCF( Discriminative Correlation Filter Tracker with Channel and Spatial Reliability) as the translate correlation filter,in order to improve the tracking of non-rectangular objects. The PCA-HOG( Principal Component Analysis of Histogram of Oriented Gradient) features are also used to improve the scale correlation filter by reducing the dimensions of HOG( Histogram of Oriented Gradient) features without information loss,and accelerate the calculation of fourier transform by using projection matrix.The evaluation of the experiments is based on VOT2018 benchmarks and video sequences. The results show that the improved algorithm performs better robustness than CSR-DCF.
引文
[1]Henriques J F,Caseiro R,Martins P,et al. High-Speed Tracking with Kernelized Correlation Filters[C]. IEEE TPAMI,2014,37(3):583-596.
[2]Martin D,Gustav H,Fahad S,et al. Accurate Scale Estimation for Robust Visual Tracking[C]. BMVC,2014.
[3]Alan L,Tomǎs V,LukǎC,et al. Discriminative Correlation Filter Tracker with Channel and Spatial Reliability[C]. CVPR,2017:4847-4856.
[4]Martin D,Gustav H,Fahad S,et al. Discriminative Scale Space Tracking[C]. TPAMI,2016,39(8):1561-1575.
[5]VOT 2018 dataset preview[EB/OL].(2018-9-13)[2018-11-20]. http:∥www. votchallenge. net/vot2018/dataset. html.
[6]The sixth Visual Object Tracking VOT2018 challenge results[EB/OL].(2018-9-13)[2018-11-20]. http:∥www. votchallenge.net/vot2018/results. html.
[7]Cehovin L,Leonardis A,Kristan M. Visual object tracking performance measures revisited[C]. IEEE Transactions on Image Processing,2016,25(3):1261-1274.
[2]Martin D,Gustav H,Fahad S,et al. Accurate Scale Estimation for Robust Visual Tracking[C]. BMVC,2014.
[3]Alan L,Tomǎs V,LukǎC,et al. Discriminative Correlation Filter Tracker with Channel and Spatial Reliability[C]. CVPR,2017:4847-4856.
[4]Martin D,Gustav H,Fahad S,et al. Discriminative Scale Space Tracking[C]. TPAMI,2016,39(8):1561-1575.
[5]VOT 2018 dataset preview[EB/OL].(2018-9-13)[2018-11-20]. http:∥www. votchallenge. net/vot2018/dataset. html.
[6]The sixth Visual Object Tracking VOT2018 challenge results[EB/OL].(2018-9-13)[2018-11-20]. http:∥www. votchallenge.net/vot2018/results. html.
[7]Cehovin L,Leonardis A,Kristan M. Visual object tracking performance measures revisited[C]. IEEE Transactions on Image Processing,2016,25(3):1261-1274.