基于多模板匹配的双模型自适应相关滤波跟踪算法

为有效提升目标跟踪的精确度和实时性，设计了基于多模板匹配的双模型自适应相关滤波跟踪算法。对多模板匹配模型与核相关滤波跟踪模型参数进行初始化处理：多模板匹配模型选取得分函数作为模板与候选样本间匹配准则，通过候选样本得分获取最佳目标，更新多模板后，通过形变多样相似性实现多模板匹配；核相关滤波跟踪模型利用所采集目标样本数据建立循环矩阵，通过训练核化岭回归分类器获取核相关滤波器，并获取响应置信图，再利用响应置信图获取下一帧图像目标位置。通过自适应融合策略获取两个模型所估计目标位置，再采用金字塔尺度估计策略估计目标尺度变化，通过不断更新各模型参数实现目标精准跟踪。实验结果表明，在目标受遮挡或旋转、光照变化等复杂环境下，该算法的中心跟踪误差均低于15 dpi，平均跟踪精确度均高于98%，且目标定位时间低于100 ms，说明该算法在跟踪精确度和实时性上具有明显的应用优势。

Dual-model adaptive correlation filter tracking algorithm based on multi-template matching

In order to effectively improve the accuracy and real-time performance of target tracking, a dual-model adaptive correlation filtering tracking algorithm based on multi-template matching is designed in this study. The parameters of the multi-template matching model and the kernel-related filtering tracking model are initialized firstly. Among them, the multi-template matching model takes the score function as the matching criterion between the template and the candidate sample, obtains the best target through the candidate sample score, and realizes the multi-template matching through the deformation and diversity similarity after updating the multi-template. The kernel correlation filter tracking model uses the collected target sample data to establish a circulant matrix, obtains the kernel correlation filter and the response confidence map by training the core ridge regression classifier. Then the target position of the next frame of image is obtained by using the response confidence map. An adaptive fusion strategy is adopted to obtain the estimated target position of the two models, and then the pyramid scale estimation strategy is employed to estimate the target scale change. Accurate target tracking is achieved by continuously updating each model parameter. The experimental results show that the center tracking error of the algorithm is lower than 15 dpi, the average tracking accuracy is higher than 98%, and the target positioning time is less than 100 ms in complex environments such as target occlusion or rotation and illumination changes. The above results indicate that the algorithm bears obvious application advantages in tracking accuracy and real-time performance.