Abstract: Moving target tracking is a hotspot in the field of computer vision. For the current moving target method, the overall information in the target frame will be considered when tracking the target, but the shape of the tracked target is variability, resulting in low robustness and accuracy of the tracking method. In this paper, using the local structure of the tracked target, the tracking target is regarded as a block particle. Under the sequence Monte Carlo method, the block particle confidence function and the motion similarity function are constructed, and the best position of the tracking target is calculated by weight voting. Based on the problem of insufficient single feature extraction based on basic tracking method, a method based on multi-feature linear fusion of HOG and color histogram is proposed. Experiments show that compared with a variety of excellent tracking methods, the proposed method achieves the lowest average tracking error of 14.83 and the target overlap rate of 0.67, which indicates that the algorithm can perform target tracking stably and accurately.

Key words: target tracking, effective blockparticle, Monte Carlo method, multi-feature fusion

摘要： 运动目标跟踪是计算机视觉领域的一个热点，针对目前运动目标方法在跟踪目标时候会考虑目标框中整体信息，但是被跟踪目标外形具有多变性，导致跟踪方法鲁棒性和精准度不高的问题。利用被跟踪目标局部结构，将跟踪目标看成块粒子，在序列蒙特卡方法下，构建块粒子置信函数和运动相似性函数，通过权重投票计算跟踪目标最佳位置。并基于基础跟踪方法单一特征提取不足的问题，提出一种基于HOG和颜色直方图多特征线性融合的方法。实验表明，与多种优秀跟踪方法相比，该方法取得最低平均跟踪误差14.83，且目标重叠率达到0.67，表明该算法可以稳定、准确地进行目标跟踪。

关键词: 目标跟踪, 有效块粒子, 蒙特卡罗方法, 多特征融合