一种结合Camshift和Kalman滤波的TLD目标跟踪算法

TLD(Tracking-Learning-Detection)跟踪最大的优点是对初始选择的目标进行不断的学习,来获取目标当前的外观特征信息。但其计算量大,当有相似目标出现、目标物被遮挡时,跟踪精确度低、效果差。Camshift算法是基于Meanshift算法形成的可连续自适应的一种算法。Camshift结合Kalman滤波可实现对目标位置的快速查找和对窗口大小的控制功能。将TLD跟踪方法的原始输出数据与改进算法的预测结果结合,再修正当前时刻的状态输出结果。对输出结果加权处理,得到目标的最终准确位置。改进算法既具有TLD算法原有的长期有效跟踪特点,又提高了对目标实时跟踪的准确性,同时对短时遮挡具有预测功能。

A TLD TARGET TRACKING ALGORITHM COMBINING CAMSHIFT AND KALMAN FILTERING

The biggest advantage of TLD (Tracking-Learning-Detection) is that the initially selected target is continuously learned to obtain the current appearance and feature information of the target. However, it has a large amount of calculation. When a similar target appears or the target object is blocked, the tracking accuracy is low and the effect is poor. Camshift algorithm is a continuously self-adaptable algorithm based on Meanshift algorithm. Combining with Kalman filtering, Camshift can quickly find the target position and control the window size. We combined the original output data of the TLD tracking method with the prediction result of the improved algorithm, corrected the state output result at the current moment, and weighted the output result to obtain the final accurate position of the target. The improved algorithm not only has the original long-term effective tracking characteristics, but also improves the accuracy of real-time tracking of the target, and has the function of predicting short-term occlusion.