多特征融合相关粒子滤波器视频目标跟踪算法-《计算机技术与发展》

文章信息/Info

Title:: Multiple Features Fusion Targets Tracking Method Based onCorrelation Particle Filter

Author(s):: ZHANG Ru-jia; YANG Xiao-jun; WANG Hai; School of Information Engineering,Chang’an University,Xi’an 710001,China

Keywords:: visual target tracking; correlation particle filter; multiple features fusion; color feature; edge feature

摘要:: 针对视频目标跟踪过程中目标形变和遮挡等一系列挑战性问题导致单一特征目标跟踪鲁棒性弱的问题, 提出了一种基于相关粒子滤波框架下的多特征融合的视频目标跟踪算法。在相关粒子滤波算法框架下,对于每个抽样粒子,首先选取色彩特征作为观测量进行滤波, 并用相关滤波器训练色彩特征,得到每个粒子的响应图; 再从得到的每个粒子响应图上选取响应值最大位置处作为第二次滤波的粒子分布位置,针对每个粒子,再次选取边缘特征作为观测量进行滤波,并用相关滤波器训练边缘特征,得到每个粒子的响应图,最后从得到的响应图中选取峰值位置,得到的位置即为目标最终预测的位置。实验结果表明,相对于现今已有的跟踪算法,该算法显示出优越的性能,在遮挡、目标形变以及复杂背景等具有挑战性因素的影响下仍然可以精确地跟踪目标,展现出更强的鲁棒性。

Abstract:: Aiming at the problem of occlusion and deformation in video target tracking, which leads to the weak robustness of single feature target tracking,a video target tracking algorithm based on hierarchical multiple features fusion under the framework of related particle filtering is proposed. Under the framework of the relevant particle filtering algorithm, the color features are selected as observations for first- level filtering. For each sampled particle, the correlation filter is used to guide the particles to the target state distribution mode,and the proposed distribution of the particle filter is constructed. It is proposed that the distribution be sampled for importance. For each particle after re-sampling,the edge feature is again selected as the observation to perform second -level filtering to obtain a posterior probability density estimate that tracks the target state, and finally determine the precise state of the target. The experiment shows that compared with the existing tracking algorithms,the proposed algorithm shows superior performance. It can still accurately track the target under the influence of challenging factors such as occlusion,target deformation and complex background,showing stronger robustness.