基于改进的随机森林TLD目标跟踪方法

针对TLD算法中采用的随机森林分类器的决策树阈值固定,不能根据目标特征随时调整,影响分类精度和时间开销的问题,引入极端随机森林的思想,提出了基于改进的随机森林TLD目标跟踪方法。该方法用Gini系数度量样本集合的混乱程度,通过比较Gini系数是否超过了给定阈值,判断叶节点何时转变成决策节点进行分裂;再结合TLD算法中的P-N学习框架和在线模型训练更新样本;最终基于改进的TLD算法完成目标跟踪。将本文方法应用于多个视频集进行目标跟踪实验,验证了算法的有效性和稳定性。

TLD Target Tracking Method based on Improved Random Forest

The fixed threshold value of the random forest classifier can not be readily adapted to the target feature of TLD (Tracking Learning Detecting) algorithm, which affects classification accuracy and time overhead issues. Aiming at this problem, an idea of extreme random forest was introduced, and a TLD target tracking method based on improved random forest is proposed. This method used Gini coefficient to measure the degree of confusion of sample sets, by deciding the Gini coefficient whether exceeded a given threshold to judge when a leaf node changed into the decision node to split. Combining P-N learning framework and online model, samples were trained and updated. Finally, the target tracking was completed based on improved TLD algorithm. The proposed method was used in many video sets for target tracking to verify its effectiveness and stability.