基于改进KCF的跟踪注册方法

针对三维注册易受环境以及目标跟踪检测算法耗时严重、精度低的影响，提出改进KCF(I KCF)的跟踪注册方法。该方法分为4步：（1)利用正则最小二乘分类器的样本训练来获取尺度核相关滤波器和位置信息；（2)搜索尺度核相关滤波器和位置输出响应最大值，完成尺度和目标位置的检测；（3)借鉴MOSSE跟踪器更新方法对模型更新；（4)采用ORB算法对目标位置特征检测并计算出注册矩阵。选取视觉跟踪基准数据集中的6组数据以及拍摄的视频序列仿真实验。仿真结果表明，当目标位置发生旋转、缩放、部分遮挡、光照和运动模糊时，I KCF在精确度、成功率以及效率上总体优于KCF、TLD、Struck和CT算法；且目标位置与OpenGL立方体注册融合度较高；基于I KCF的AR系统具有较好的实时性、稳定性和鲁棒性。

A tracking and registration method based on improved KCF

Since 3D registration is easily affected by the environment and target tracking and detection algorithms are time consuming with low precision, we propose a tracking and registration method based on an improved kernerlized correlation filter (I KCF). The method includes four steps: (1) utilizing the regularized least squares classifier for sample training to obtain kernel correlation filter and position information; (2) searching scale kernel correlation filter and the maximum of position output to achieve the detection of the scale and position; (3) updating the model by referring to the MOSSE tracker; (4) adopting the oriented FAST and rotated BRIEF (ORB) to do feature extraction and matching, and then calculate the registration matrix. We utilize 6 sets of data in the Visual Tracker Benchmark datasets and video sequence to simulate. The results show that the I KCF generally outperforms the KCF, tracking learning detection (TLD), structured output tracking with kernel (Struck) and compressive tracking (CT) in precision, success rate and efficiency when rotation, scale variation, partial occlusion, illumination or motion blur occurs. Besides, the target position is highly aligned with OpenGL cube registration, and the augmented reality (AR) system based on I KCF is more real time, stable and robust.