多目标跟踪是计算机视觉领域里一个重要的研究课题。它在军事视觉制导、交通管制、校园和小区的视频监控等领域中有着广泛的应用。但同时，跟踪过程中背景环境的复杂性，跟踪精度与速度的平衡性以及遮挡等问题，都使得完成精准快速的有效跟踪是一个要继续面临的挑战。 本文首先研究了Meanshift算法在多目标跟踪中的应用，通过实验得到，该方法在使用单一特征跟踪目标时，在背景对目标产生强烈干扰的情况下，会发生跟踪丢失的现象。在此基础上，本文对算法进行了改进。以目标和背景的区分度作为评判准则，认为区分度最大的特征是把目标从背景中分割出来的最佳特征，分别计算RGB颜色特征、LBP纹理特征以及Canny边缘特征下目标和背景的区分度，并通过粒子群优化算法计算每种特征的权值也即对分割目标的贡献度，之后对各特征进行加权获得综合特征，以该特征作为Meanshift算法中迭代计算目标位置特征。改进后的方法包含了多种特征信息，且每次都能找到最优的权值分配，依靠特征之间的互补性，使得在复杂背景下跟踪目标更精确。 其次，针对上述方法使用大量的信息导致不能满足跟踪实时性的问题，使用串行特征迭代的策略，在Menashift算法的第一次迭代中使用融合后的综合特征来计算目标模型和候选目标模型，计算一次迭代后目标所在的位置，从第二次以后，从融合之前的各特征中选出权值最大的一种特征，也即分割目标和背景最可靠的特征作为迭代特征，迭代计算直到确定目标在当前帧中的位置。因为在除第一次之外的每次迭代中都只使用了一种特征，所以和之前使用综合特征迭代相比，大大减小了计算量，并且选择了可靠性最高的一种特征来判断，在提高了跟踪精准度的同时，也保证了算法的实时性。 最后，在上述方法的基础上，引入了卡尔曼滤波的方法对其进行优化，分别使用上述方法和卡尔曼滤波器进行目标位置的计算，并设置阈值与两种算法计算得到的位置差进行比较，位置差小于阈值时，取两个位置的中值作为当前帧目标的位置，否则，选择与目标模板相似度大的位置做为当前帧的目标位置。由模板相似度和滤波器残差判断出目标发生遮挡时，使用卡尔曼滤波预测目标位置，遮挡结束时，继续使用遮挡前的方法进行跟踪，可以有效地跟踪被遮挡目标。

Multi-target tracking is an important research topic in the field of computer vision. It has broad applications in the military visual guidance, traffic control, campus and community of video surveillance and other fields. But at the same time, during the tracking the diversity of goals pattern, the complexity of the background environment, the balance of accuracy and speed and the occlusion make that completing tracking fast and effectively is still a challenge which we have to face up. First, the essay researches the application of the traditional Meanshift algorithm in multi target tracking, through the experiments some completion are found that the targets would lost in the background which make strong interference to the targets. Based on this, the traditional algorithm has been improved. That, the discrimination between the target and the background is regards as the evaluation criteria, the feature which has the max discrimination is thought of the best feature to segmentation the targets from the background in this essay . Calculating the discrimination between the targets and background of the RGB color features, LBP texture feature and Canny edge features respectively, and then calculating the weights of each feature by Particle swarm optimization algorithm, and the weights means the contribution various features to segmented the targets from the background, and access to fusion feature through weighting each features. Last, using the fusion features to calculate the targets’ position iteratively by Meanshift algorithm to complete the targets tracking. The improved method includes a variety of characteristic information, and could find the best weights every time, depend on the complementarity between features; the traditional method which based on one single feature under the complex background is improved. Then, aimed at the problem that a lot of information used in the above method leads that the method could not meet the real-time, the strategy of serial feature is used into iteration. In the first iteration of Meanshift algorithm the fusion feature is used to compute the target model and candidate target model, and through this ,the positions of targets are obtained, but from the second time, the best feature which has the biggest weight is selected from the above features, and this feature is the most reliable feature which is regards as iterative feature to distinguish the targets and background, and use the feature to determine the target position in the current frame iteratively. Through this improved method , the amount of calculation are greatly reduced, because in each of iteration in addition to the first just one single feature is used, but not use fusion features in every iterations like before, and select the most reliable feature to judge the targets, the accuracy and real-time of the algorithm in targets tracking is improved. Finally, the Kalman filter method is used to optimize the above methods and the position of the targets are calculated by the above method and Kalman filter separately. And set a threshold to compare with position difference by the two methods. If the threshold is bigger, set the middle of the two position as the targets’ position, else select which has the large similarity with target template. By the template similarity and filter residuals judged that the target is occluded, if occluded, using Kalman filter to calculate the position, when targets appear, use the previous method to track continuously.