一种多特征自适应学习机制的目标跟踪算法

Object Tracking Algorithm Based on Multi-feature Adaptive Learning Mechanism

当前,目标跟踪技术日益成熟,但是受应用场景等因素的影响,比如当目标发生明显的形变、旋转、尺度变换、光照变换、运动模糊等情况传统的目标跟踪算法的准确度会大打折扣。为提高传统相关滤波(KCF)在目标跟踪上的准确度,文中在KCF目标跟踪的框架基础上,首先,独立训练HOG+Gray和CN颜色特征两个滤波器并输出对应响应图,然后通过自适应加权的方式融合两者的输出响应图;其次,在尺度变化方面,通过建立目标尺度池来求取目标最佳尺度,并且融入反向尺度检测机制;最后,在目标快速运动和遮挡问题上,通过自适应调整模型的学习率更新模型。实验部分,采用OTB-100目标跟踪数据库进行评测。将结果与常见几种跟踪算法进行对比,在平均跟踪精度与速度上优于其他算法。

At present,the object tracking technology is becoming more and more mature,but the accuracy of traditional track⁃ing algorithms will be greatly discounted when the target has obvious deformation,rotation,scale transformation,illumination trans⁃formation,motion blurring and so on.Although the traditional correlation filter tracking algorithm(KCF)has better tracking speed and tracking accuracy in OTB-100,it still has some shortcomings,such as a single feature and inability to scale changes.In order to improve the accuracy of traditional correlation filter(KCF),based on the framework of KCF target tracking,firstly,two filters of HOG+Gray and CN color feature are independently trained and corresponding response maps are output,then the output response maps of both filters are fused by adaptive weighting method.Secondly,in the aspect of scale transformation,the target scale pool is established to get the best scale and integrates the detection mechanism of reverse scale.Finally,the learning rate of the adaptive ad⁃justment model is used to update the model for fast target movement and occlusion.In the experimental part,the OTB-100 is used for evaluation.Comparing the results with several common tracking algorithms,it is superior to other algorithms in average tracking accuracy and speed.