基于改进粒子滤波的目标跟踪方法研究

发布时间：2018-07-14 11:58

【摘要】：运动目标跟踪是计算机视觉领域中的一个十分重要的研究方向,随着对运动目标跟踪技术的研究工作得到广泛开展,运动目标跟踪技术得到了快速发展。与此同时,人们对于运动目标跟踪技术的要求也是与日俱增,如何对各种复杂场景中的运动目标进行准确和稳定的跟踪一直是运动目标跟踪领域的难点。本课题的研究工作主要包括以下几个方面:1、针对多区域采样目标跟踪方法容易出现的区域多样性丧失、跟踪精度下降和跟踪不稳定等问题,通过引入区域优化权值及改进子区域重采样方法,给出基于优化权值的多区域采样目标跟踪算法。该方法利用区域优化权值优化各个子区域的区域置信度适当增加低置信度区域在重采样阶段所分配到的粒子数量,在保证粒子根据区域置信度大小有效分配的前提下,抑制了区域多样性丧失现象发生。该方法在子区域内引入粒子权重优化权值并设定重采样阈值,缓解粒子贫化充分利用有效粒子信息。实验结果表明,该方法能有效提高目标跟踪精度,改善目标跟踪稳定性。2、针对传统粒子滤波算法粒子重采样产生的粒子贫化现象及单一特征目标跟踪算法鲁棒性较差的问题,给出一种基于信息保留的自适应多特征融合目标跟踪算法。该算法的信息保留策略在粒子重采样阶段通过优化粒子权重值分布来适当提高小权重粒子的权重并改进了粒子重采样方法,有效抑制了粒子贫化现象,保留更多粒子信息。根据环境变化对特征有效性的影响及不同特征对目标的贡献度,自适应调节多特征模型中各特征分量的权重。实验结果表明,该算法能有效应对目标形变、目标部分遮挡、背景相似物体干扰等复杂情况,具有良好的跟踪精度和鲁棒性。3、针对运动目标跟踪过程中容易受到复杂环境及目标遮挡影响的问题,给出一种结合全局特征融合和局部均值漂移的目标跟踪算法。该算法将目标区域划分为多个子区域,将粒子滤波方法和均值漂移方法分别应用于目标的全局区域和局部子区域的跟踪中。采用改进的粒子滤波方法并融合颜色和FDF特征进行目标全局区域跟踪,利用均值漂移算法并融合颜色和纹理特征进行目标子区域跟踪。该算法通过目标受到遮挡程度来自适应调节全局信息和局部信息在目标跟踪中的贡献,提高了目标跟踪算法应对遮挡场景的适应性,融合多特征改善了目标跟踪算法对复杂跟踪场景的鲁棒性。实验结果表明,该算法能有效应对目标形变、目标遮挡和复杂背景干扰等影响,具有良好的跟踪稳定性和精确度。
[Abstract]:Moving target tracking is a very important research direction in the field of computer vision. With the development of moving target tracking technology, moving target tracking technology has been developed rapidly. At the same time, the requirement of moving target tracking technology is increasing day by day. How to track moving targets accurately and stably in various complex scenes is always a difficult point in the field of moving target tracking. The research work of this subject mainly includes the following several aspects: 1, aiming at the problems such as the loss of regional diversity, the decline of tracking accuracy and the instability of tracking, which are easy to appear in the multi-region sampling target tracking method. By introducing regional optimization weights and improved subregion resampling method, a multi-region sampling target tracking algorithm based on optimal weights is presented. In this method, the regional confidence of each sub-region is optimized by using the regional optimization weights to increase the number of particles assigned to the low-confidence region in the resampling stage, while ensuring the effective distribution of the particles according to the confidence level of the region. The loss of regional diversity was restrained. In this method, the particle weight optimization value is introduced into the sub-region and the resampling threshold is set so as to reduce particle dilution and make full use of the effective particle information. The experimental results show that the proposed method can effectively improve the tracking accuracy and the stability of target tracking. Aiming at the problem of particle dilution caused by particle resampling in traditional particle filter algorithm and the poor robustness of single feature target tracking algorithm, the proposed method can effectively improve the tracking accuracy. An adaptive multi-feature fusion target tracking algorithm based on information reservation is presented. In the phase of particle resampling, the information retention strategy of the algorithm can appropriately increase the weight of small weight particles and improve the method of particle resampling by optimizing the distribution of particle weight values, which can effectively suppress the phenomenon of particle dilution and retain more particle information. According to the effect of environmental change on feature validity and the contribution of different features to the target, the weight of each feature component in the multi-feature model is adjusted adaptively. The experimental results show that the algorithm can effectively deal with complex situations such as target deformation, partial occlusion of the target, background similar object interference and so on. It has good tracking accuracy and robustness. Aiming at the problem that moving target tracking is easily affected by complex environment and target occlusion, a target tracking algorithm combining global feature fusion and local mean shift is proposed. The algorithm divides the target region into several subregions and applies the particle filter method and the mean shift method to the tracking of the global region and the local subregion of the target, respectively. The improved particle filter method is used to track the target global region by combining color and FDF features, and the mean shift algorithm is used to track the target subregion with the fusion of color and texture features. The algorithm adaptively adjusts the contribution of global and local information in target tracking by the degree of occlusion, and improves the adaptability of the target tracking algorithm to the occlusion scene. Fusion of multiple features improves the robustness of the target tracking algorithm to complex tracking scenarios. Experimental results show that the algorithm can effectively deal with the effects of target deformation, target occlusion and complex background interference, and has good tracking stability and accuracy.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】