基于多特征融合与背景目标双加权的行人跟踪  

PEDESTRIAN TRACKING BASED ON MULTI-FEATURE FUSION AND BACKGROUND-TARGET DOUBLE WEIGHTING

作　　者：卜言生 贺俊吉[1] Bu Yansheng;He Junji(Logistics Engineering College,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学物流工程学院,上海201306

出　　处：《计算机应用与软件》2020年第10期225-231,238,共8页Computer Applications and Software

摘　　要：传统Mean Shift跟踪算法表观行人时的特征单一且忽略背景因素,难以在复杂环境下对目标行人进行有效跟踪。针对此问题,采用RGB颜色特征与LBP纹理特征对目标行人进行模型描述,并由S型函数自适应融合两类特征。考虑到目标行人自身的特征以及背景因素,对目标行人进行背景目标双加权,从而增强模型的描述能力;为削弱外部环境(如:光照变化、遮挡等)对模型的影响,采用特征更新选择函数,用于跟踪过程中模型分量的选择性更新。通过实验和定量分析表明:该算法能在光照变化和短时遮挡等情况下实现目标行人的有效跟踪,改善了传统Mean Shift算法中模型描述不准确的局限性。The traditional Mean Shift algorithm only adopts a single feature to describe the target pedestrian and ignores background,making it difficult to effectively track target pedestrian in complex environments.To solve this problem,the target pedestrian is modeled by color features and LBP texture features,and two types of features are adaptively fused by S-type function.Considering the characteristics of the target pedestrian and the background,the target pedestrian was double-weighted to enhance the description ability of the model.To weaken the influence of the external environment(such as:illumination changes,occlusion,etc.)on the model,the feature update selection function was used to track the selective update of the model components in the process.The experimental results and quantitative analysis show that the algorithm can effectively track the target pedestrian under illumination changes and short-term occlusion,and improve the limitations of inaccurate model description in the traditional Mean Shift algorithm.

关 键 词：行人跟踪 特征自适应融合 背景目标双加权 选择性更新 Mean shift 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]