【作者】 吴刚；

【导师】 唐振民；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2014， 博士

【摘要】 现代交通运输智能化涉及车辆自身的智能化与道路交通智能监控。智能车辆采用视觉系统针对周围车辆进行跟踪,有助于对潜在接近或危险行驶的车辆进行合理规避；道路交通智能监控环境中,针对车道上的行驶车辆进行目标跟踪,有助于提取交通流速度、车道的占有率与瞬时车速等交通信息量。符合目标状态非线性、噪声分布非高斯的粒子滤波算法由于既适用静止视觉平台,同时又适合移动视觉平台,成为目前跟踪算法的一个较优选择。在图像受随机噪声扰动、车辆受遮挡、车速变化等场景下,传统车辆跟踪算法的成功率与处理速度均难以完全匹配智能车视觉系统的要求；对于道路交通智能监控,基于图像的车辆跟踪算法也存在不能完全适应光照强度变化、阴影、雨雪雾天等问题。作为传统预测跟踪方法,粒子滤波算法自身缺乏智能性应对复杂场景车辆跟踪的有效解决手段。针对上述复杂车辆跟踪场景,论文侧重在融合粒子滤波与增量学习方面,研究进一步提高粒子滤波跟踪算法稳定性、实时性与可靠性的解决方法。本论文主要研究复杂场景中的运动车辆跟踪方法,研究成果与取得的创新点体现在以下主要方面：(1)为解决传统粒子滤波序列重要性重采样(Sequential Importance Resampling, SIR)算法因引进简化的建议分布,引起粒子权值近似计算从而导致跟踪稳定性欠佳的问题,利用相邻帧间信息的强关联性,使用前帧先验与系统状态预测信息共同采样粒子以校正预测粒子点信息,提出一种引入前帧加权采样的粒子滤波跟踪算法。基于标准视频Car11的车辆跟踪实验表明：对道路灯光扰动下车速发生变化的车辆进行跟踪,所提算法提高粒子滤波SIR算法跟踪的稳定性。为进一步降低重采样误差并提高粒子滤波的采样效率,既通过引入当前粒子集权值的残差信息来构建合理的累积分布函数,又通过分层的手段来获取有序的随机数集合,在多项式重采样的基础上提出残差信息分层重采样。针对标准测试视频,对比嵌入4种重采样的粒子滤波SIR跟踪算法的实验数据表明：在跟踪误差与收敛粒子数方面,嵌入残差信息分层重采样的粒子滤波算法性能为最优。(2)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,基于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基于表观模型的车辆跟踪方法,不需预先训练车辆图像、且不需假定目标车辆的子空间均值固定,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。在粒子数选取300个、每5帧进行一次IPCA增量学习的情况下,基于Car4等3组标准视频的跟踪实验表明：对比P.Hall-IPCA粒子滤波与D.Ross-IPCA粒子滤波方法,所提AM-IPCA粒子滤波方法将车辆跟踪成功率分别由82.7%-92.3%、92.1%-95.2%提升至95.1%-96.4%。(3)长序列图像跟踪场景中,车辆跟踪过程经常受跟踪区域形变、光照强度变化等现场强噪声干扰,为解决跟踪算法的跟踪窗口易产生形变与漂移的难点问题,利用群空间中仿射群组受扰动后的形不变性,将系统状态变量映射到李群空间进行处理,同时采用IPCA增量学习并更新目标特征子空间。在利用粒子滤波算法采样粒子时,通过引入观测量以提高粒子权值计算的准确性。所提跟踪器可以在现场强噪声的扰动下长时段稳健地跟踪车辆,跟踪窗口自适应地调整大小与角度以不断适应车辆姿态和距离远近的持续变化。基于Dtneu_schnee等4个标准视频的跟踪实验表明：对比同类别跟踪器VTD、IVT和Kwon2010,所提自适应跟踪器将跟踪成功率分别由91.3%～95.7%、82.9%-94.2%、94.6%-96.7%提升N96.1%～97.7%。(4)针对车辆在经历遮挡、运动状态变化及雨雪噪声扰动下的跟踪难点问题,并非一味地采用预测跟踪算法对车辆进行长时段跟踪,结合粒子滤波、Online Boosting在线分类器及其增量学习、样本相似度与半监督学习提出SSOBI粒子滤波跟踪算法,利用当前时刻观测数据、相邻帧与系统状态预测信息共同增强粒子滤波的粒子采样合理性,避免粒子滤波的建议分布过于依赖系统状态预测而处于一定的局部最优范围之内；粒子滤波合理的建议分布则有效地缩小在线增量学习样本的检测范围,加快Online Boosting的学习并有效解决增量学习过程中存在的自学习问题。针对Dtneu_winter等测试视频的跟踪实验表明：车辆面临部分遮挡时,对比Online Boosting、半监督Online Boosting算法,SSOBI粒子滤波算法将跟踪成功率分别由66.5%～73.3%、76.4%～82.5%提升至88.5%-90.0%。更多还原

【Abstract】 Intelligentized trend of modern traffic transportation relates to vehicle intelligence in itself and intelligent traffic surveillance on roadway. Using vision system to track nearby vehicles helps intelligent vehicle to reasonably evade potential near or dangerous driving vehicles. On the surroundings of intelligent traffic surveillance, tracking moving vehicles on roadway helps to extract traffic informations such as traffic flow speed, occupancy rate of lanes, instantaneous vehicle speed, etc.. Particle filtering algorithm accords with the demands of nonlinear target state and non-Gaussian noise distribution, which is suitable not only for stationary vision platform but also for moving vision platform. The above advantages make it an optimum option on current tracking algorithms. On the scenes such as stochastic noises disturbing on images, vehicle blocked and vehicle speed changing, etc., the success rate and processing speed of traditional vehicle tracking algorithms are difficult to completely match the demands of vision system on intelligent vehicle. On the aspect of intelligent traffic surveillance on roadway, image-based vehicle tracking algorithms are also difficult to completely adapt to these scenes such as illumination change, shadow and rain-snow-fog weather, etc.. As the traditional predictive tracking method, particle filter algorithm itself lacks of sufficient intelligent means to dealing with vehicle tracking on complicated scenes. On above complicated scenes about vehicle tracking, the thesis emphasizes particularly on the aspect of integrating particle filter with incremental learning, researches solutions on further improving stability, realtime performance and reliability about particle filter tracking algorithm.This thesis mainly focuses on moving vehicle tracking methods on complicated scenes, main researchful work and contributions of the thesis are as follows:(1) To resolve the issue of traditional particle filter SIR algorithm which tracking stability is not good enough caused by approximate calculation of particles weights attributing to import simplified proposal distribution, based on the strong relevancy within sequential frames, particle filter tracking algorithm importing weighted sampling on pre-frames is proposed in this thesis, sampling particles on pre-frames aprior information and system state predictive information to correct information of predictable particles. The tracking results on standard testing video Carll demonstrate that the proposed algorithm can steadily track vehicle which speed is variational on road light existing scene, compared with particle filter SIR algorithm. For further decreasing errors on resampling and improving efficiency of sampling strategy on particle filter, stratified resampling method imported by residual information is proposed based on multinomial resampling method. Reasonable accumulative distributing function is constructed by importing residuals of weights on associated particles. Synchronously, sequential stochastic numbers are gradually produced by stratifing on stochastic muster. Particle filter SIR algorithms embedded by four kinds of resampling methods are tested on standard videos. Based on testing data of tracking errors and convergent particles, the tracking data indicates that particle filter tracking method embedded by the proposed resampling is best among these tracking methods.(2) 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., based on autocorrelation matrix, incremental learning on IPCA and particle filter algorithm, one kind of vehicle tracking methods using appearance model is proposed, not relying on training images of vehicle in advance and not assuming the subspace mean of vehicle is fixed. When beginning at original tracking time, the proposed method can timely learn the characteristic subspace images of vehicle, using autocorrelation matrix and eigen value decomposition. Based on IPCA incremental learning, likelihood probability density is computed on subspace mean and eigenvector, increasing computational precision on weights of particles on particle filter algorithm. The number of particles is assigned as 300 and IPCA incremental learning is updated every 5 frames in this experiment. The tracking results on three standard testing videos involving Car4 demonstrate that success tracking rate of the proposed tracking method is raised to 95.1～96.4 percent, compared with 82.7～92.3 percent of P.Hall-IPCA particle filter and 92.1～95.2 percent of D.Ross-IPCA particle filter.(3) On the tracking scene involving long sequential frames, vehicle tracking process is always subjected to severe disturbances on the scenes such as object deformed, illumination change and stochastic noises, etc.. To solve the problems of distortion and excursion with window of tracking algorithm, the systemic state variables are projected to Lie group space to dealing with in this thesis, attributing to affine group invariability on disturbances. Simultaneously, the incremental learning algorithm on IPCA is used for incrementally learning and updating characteristic subspaces of the target. Based on importing measurement vector in computing weights of sampling particles on particle filtering algorithm, the method is presented to enhance computational precision about weights. Tracking window of the proposed self-adjusting tracker is not deformed by noises, and can adjust appropriate size and angle to adapting changes continuously induced by pose and distance on target. The tracking experiments based on four standard videos involving Dtneu_schnee demonstrate that success tracking rate of the proposed self-adjusting tracker is raised to 96.1～97.7 percent, compared with 91.3～95.7 percent of tracker VTD,82.9～94.2 percent of tracker IVT and 94.6～96.7 percent of tracker Kwon2010.(4) Aiming at the difficulties on tracking moving vehicle when vehicle blocked, varying on movement state, existing noise circumstance such as rain and snow, differed from unchanging prediction algorithm about tracking on vehicle over a long period of time, the SSOBI particle filter tracking algorithm is proposed, uniting these elements such as particle filter, online boosting and incremental learning, sample similarity and semi-supervised learning. Current observational data, information on adjacent frames and predictive information on systemic state are used to boost up rationality on sampling particles in particle filter, preventing proposal distribution local optimization duing to severely depending on systemic state prediction. Attributing to reasonable proposal distribution on particle filter, detection range of online incremental learning is effectively reduced, speeding up the learning and effectively solving the self-learning problem of online boosting on incremental learning process. When target vehicle is partially blocked, the tracking experiments on testing..videos involving Dtneu_winter demonstrate that success tracking rate of the proposed SSOBI particle filter algorithm is raised to 88.5-90.0 percent, compared with 66.5～73.3 percent of online boosting algorithm and 76.4-82.5 percent of semi-supervised online boosting algorithm.更多还原