基于改进 KCF 算法和多特征融合的车辆跟踪研究

针对目前车辆跟踪研究算法中，核相关滤波算法(KCF)在复杂背景下存在特征提取单一以及尺度无法自适应的不足，本文提出一种多特征融合的尺度自适应算法。该算法以颜色直方图信息作为颜色特征，将具有更多语义信息的高层卷积特征和拥有较高分辨率的底层卷积特征作为深度特征，并与颜色特征进行自适应特征融合。然后，采用上下文图像对目标背景信息进行约束优化，并通过平均峰值相关能量检测衡量响应置信度，最后利用高置信度的跟踪结果来避免模型易受干扰的问题。通过在OTB100数据集上的实验表明，本文算法的精度分别比其他的主流跟踪算法Staple、SAMF、KCF、TLD、DSST和CSK高出4.9%, 5.7%, 10.2%, 10.3%, 23.4%, 29.7%。

Research on vehicle tracking based on improved KCF algorithm and multi-feature fusion

In view of the current vehicle tracking research algorithms, the Kernel correlation filtering algorithm ( KCF) has the shortcomings of single feature extraction and the inability to adapt the scale in a complex background, this paper proposes a multi-feature fusion scale adaptive algorithm. The algorithm uses color histogram information as color features, takes high-level convolution features with more semantic information and low-level convolution features with higher resolution as depth features, and performs adaptive feature fusion with color features. Then, the context image is used to constrain and optimize the target background information, and the response confidence is measured by the average peak correlation energy detection, and finally the high-confidence tracking result is used to avoid the problem of the model being vulnerable to interference. In addition, in order to achieve high target tracking accuracy, the algorithm in this paper uses a hierarchical model update strategy to update the extracted features. Experiments on the OTB100 data set show that the accuracy of the algorithm in this paper is better than other mainstream tracking algorithms Staple, SAMF, KCF, TLD, DSST and CSK are 4. 9%, 5. 7%, 10. 2%, 10. 3%, 23. 4%, 29. 7% higher.