融合增量主成分分析与粒子滤波的车辆表观模型跟踪  

Vehicle Appearance Model Tracker Integrating Incremental Principal Component Analysis and Particle Filter

作　　者：吴刚 曾晓勤[3] 苏守宝[1] 王池社 

机构地区：[1]金陵科技学院计算机工程学院,江苏南京211169 [2]金陵科技学院大数据研究院,江苏南京211169 [3]河海大学计算机与信息学院,江苏南京210098

出　　处：《南京师大学报（自然科学版）》2017年第1期33-38,共6页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：金陵科技学院高层次人才科研启动项目(jit-rcyj-201508);国家自然科学基金项目(61375121);国家自然科学基金项目(61305011);南京市经信委项目(交通大数据公共服务平台);南京市科委重大项目(大数据驱动下的大型客运枢纽监控预警与应急处置)

摘　　要：针对车辆运动方向持续变化、目标车辆距离远近变化、光照强度变化等场景下,稳定且实时性地跟踪车辆的难点问题,融合自相关矩阵增量主成分分析(Incremental Principal Component Analysis,IPCA)增量学习与粒子滤波算法的基础上,提出一种新的基于表观模型(Appearance Model,AM)的车辆跟踪方法,从跟踪初始利用自相关矩阵与特征值分解构建车辆的子空间图像,通过IPCA增量学习后的子空间均值、特征向量基共同参与似然概率密度的计算,提高粒子滤波算法粒子权值计算的精度.标准视频的跟踪实验表明:对比P.Hall-IPCA与D.Ross-IPCA表观模型跟踪方法,所提AM-IPCA车辆跟踪方法将跟踪成功率分别由82.7%~92.3%、92.1%~95.2%提升至95.1%~96.4%.Aiming at the difficulties on stably and timely tracking vehicle on the scenes such as volatile moving direction, varying pose and distance, illumination change, etc., integrating autocorrelation matrix, incremental learning on IPCA and particle filter algorithm, new kind of vehicle tracking methods using appearance model is proposed. When beginning at original tracking time, the proposed method can timely learn the characteristic subspace images of vehicle, using autocor- relation matrix and eigen value decomposition. Based on IPCA incremental learning, likelihood probability density is com- puted on subspace mean and eigenveetor ,increasing computational precision on weights of particles on particle filter algo- rithm. The tracking results demonstrate that success tracking rate of the proposed AM-IPCA vehicle tracking method is raised to 95.1% ～ 96.4%, compared with 82.7% ～ 92.3% of P. Hall-IPCA and 92.1% ～ 95.2% of D. Ross-IPCA appear- ance model tracking method.

关 键 词：车辆跟踪 表观模型 自相关矩阵 增量学习 粒子滤波 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]