针对TLD算法的特征点无法有效表述目标问题, 提出了一种基于角点增强改进的TLD目标跟踪算法。改进算法在跟踪模块加入了对目标表述能力更强, 具有光照不敏感性和旋转不变性的Shi-Tomas角点作为跟踪特征点。跟踪器运行时, 在角点经光流法跟踪和双向误差检测后, 利用剩余的稳定角点定位目标窗口。对照结果表明, 改进算法在面对目标抖动和形变时可以稳定跟踪; 有效抑制因跟踪平滑点造成的漂移现象; 提高了跟踪的稳定性。针对TLD算法跟踪过程中因在线模板积累造成的计算量持续增大、实时性持续降低的问题, 提出了一种依据相似度中值的模板判断删除机制。该删除机制在模板积累到设定阈值时运行, 根据模板与当前目标的相似度, 删除不再具备代表性的模板; 调整模板空间并更新模板数目。实验表明, 该删除机制在应对模板更新快、持续时间长的跟踪情景时有效降低算法计算量, 实时性可提高约20%。

A new object tracking algorithm has been proposed which is based on the corner reinforced for the problem of the feature points can’t represent the target effectively in the TLD. The Shi-Tomas corner point has been added into the existing tracking points as the feature points, which is insensitive to the light change and invariant to rotation. When tracker is running, the target will be located by the reserved corner which is tracked by the optical flow and filtered by the error detector. The experiments show that the algorithm can track the target steadily when faced with the target shaking and appearance changing. And the algorithm can effectively prevent the drift tracking caused by the tracking smooth pixels and improve the tracking stability. In addition, a model deleting after judging system has been proposed for solving the problem of calculated amount increasing and real time decreasing caused by the online model accumulating during the TLD tracking. The system will delete the unrepresentative models according to the similarity of the target and the online model when the model quantity reaches to a certain amount. The experiments show that the system can decrease the computational complexity effectively, and increase the real time by 20% especially handling with the rapid model updating and longtime tracking circumstances.