车载视频下改进的核相关滤波跟踪算法

Improved Kernel Correlation Filtering Tracking for Vehicle Video

针对相关滤波跟踪算法在车载视频下由于环境复杂及目标尺度变化等情况下容易跟踪失败的问题,该文提出一种基于背景信息的尺度自适应相关滤波跟踪算法。首先利用背景感知相关滤波跟踪器融合方向梯度直方图特征预测目标下一帧位置,然后根据预测位置选取图像块进行检测,最后结合动态尺度比例金字塔模型对目标进行尺度估计。实验选取了KITTI数据库中23段车载视频和标注国内的4段车载视频进行测试,实验结果表明,该算法能有效降低车载环境的复杂背景、目标尺度变化等因素干扰,整体性能优于KCF,DSST,SAMF,SATPLE等主流相关滤波算法,对车载环境下复杂背景和尺度变化的目标跟踪具有鲁棒性。

For videos captured by in-car cameras, the filter-based tracking is a challenging task due to complex environments and mutable object scales. A scale adaptive tracking filter is proposed based on the background information. Firstly, the relative motion of each object is estimated by extracting features from gradient histograms between frames. Then, the object location on the next frame is determined and utilized to delimit an image block. Finally, the object scale is obtained through dynamic scaling pyramid model within image block. The proposed algorithm is examined by 27 in-car videos including 23 KITTI videos and 4 domestic videos. In experiments, the proposed algorithm suppresses effectively the interferences of environments and objects. It achieves more accurate and more robust object tracking than several popular benchmarks including KCF, DSST, SAMF, SATPLE.