改进的KCF实时目标跟踪算法

Improved KCF real-time target tracking algorithm

提出了一种改进的核相关滤波(KCF)实时目标跟踪算法.增加初始化跟踪范围自适应的策略,同时在跟踪过程中采用动态学习率.利用初始化跟踪范围自适应方法对跟踪目标响应趋势进行快速判断,若响应值快速衰减,则扩大跟踪范围.动态学习率的设定也与目标响应强度密切相关,当目标响应强度较高时,目标学习率较高,增加模型的鲁棒性;当目标响应值低时,模型学习率降低,减少模型保护的噪声信息.最后介绍了DSP6455定点单核定点处理器上的算法移植工作,实测耗时为23 ms,满足实时性要求.

An improved kernel correlation filter(KCF) algorithm,real-time target tracking algorithm was proposed.The adaptive strategy of initial tracking range was added.At the same time,the dynamic learning rate was used in the tracking process.The adaptive method of initial tracking range was used to quickly judge the response trend of initial tracking target.If the response value decreases rapidly,the tracking range would be expanded dynamically.The setting of learning rate was also closely related to the target response intensity.When the target response intensity was relatively high,the target learning rate increases the robustness of the model.When the target response value was low,the model learning rate decreases and the noise information of model protection was reduced.Finally,the algorithm on DSP6455 fixed-point single-core fixed-point processor is introduced.The migration work takes 23 ms for the final measured algorithm,which meets the real-time requirement.