交通监控系统中目标跟踪与行为识别研究

【作者】 吕斌；

【导师】 夏利民；

【作者基本信息】 中南大学 ， 交通信息工程及控制， 2010， 硕士

【摘要】 智能交通监控系统能够对交通事件进行自动化检测,对行人或车辆进行智能化监视,更能适应实际应用的需要。论文主要对智能交通监控系统中的目标检测、目标跟踪、以及目标行为分析理解三个环节中存在的关键问题进行深入研究,并提出新的解决方法,主要工作体现在以下几个方面：(1)针对当前大多利用单一模型进行目标检测存在的问题,比如高误检率,光照敏感,动态场景鲁棒性差等问题,提出了一种混合运动检测模型,将对光照变化不敏感的目标检测模型和对动态场景变化跟踪能力快的运动检测模型融合,利用融合策略消除检测过程中的漏检和误检。最后提出利用快速运动目标检测法减少该模型的计算量,加上被融合的两种模型都有较好的实时性特点,使得混合模型仍然具备一定的实时性。(2)研究了跟踪过程中的目标描述,提出一种基于多特征选择的运动目标跟踪算法。将RankBoos与AdaBoost组合,构建混合boosting算法,根据目标信息和背景信息选择特征,建立特征排序分类器,并在跟踪的过程中不断自适应更新。采用卡尔曼滤波对目标区域进行粗预测,然后利用排序分类器结合Mean-shift算法完成目标的精确跟踪。该算法可以根据不同的目标和背景信息,自适应的进行特征选择,对于克服场景中存在光照、干扰、遮挡等问题是非常有利的。(3)提出了一种基于轨迹分析的运动行为识别方法。通过采用聚类的方法对跟踪得到的轨迹进行行为模式学习得到运动模式的轨迹参考序列。然后将轨迹视为时间序列,利用动态时间归正(DTW)技术对时间序列长度没有限制的特性,将DTW与K近邻算法结合用于待识别轨迹与参考序列模板轨迹的匹配,匹配过程中,采用DTW下界函数剔除大量不相似轨迹,以加快匹配速度,进而识别目标的运动状态。实验结果表明,本文的目标检测、跟踪算法可以对目标进行有效的检测和稳定跟踪,基于轨迹分析的运动行为识别方法在十字路口行人的左转,右转,前行,U型转达到了较高的识别率。更多还原

【Abstract】 Intelligent traffic surveillance system with the characteristics of automatic and intelligent, has the ability of detecting traffic incidents, monitoring the pedestrians and vehicles in the traffic sence,can adapt to the needs of practical application. This paper focuses on object detection, object tracking, and object behavior analysis in the intelligent traffic surveillance system, then delves into the key problems in these three technologies,proposes new solutions. The paper works in the following aspects:(1) Most of current methods use a single model for the object detection exist many problems, such as high error rate, light sensitivity, poor robustness in the dynamic scenes,because of this,the paper presents a hybrid model of motion detection base on a certain blending rules, we integrate object detection model which is not sensitive to light changes and the other target detection model which can track scene changes quickly into a mixed-target detection model.The blending strategies are help for eliminating missed and false detections. Finally, the fast moving target detection method is used to reduce the computation of this model, together with the two models which have been integrated well both have simple computation procedure, the hybrid model still runs in real-time.(2) In the tracking, the paper mainly studies the object description in the tracking process, proposes a tracking algorithm base on multi-feature selection. We combinate RankBoost with AdaBoost to construct hybrid-boosting algorithm, then use hybrid-boost and the informations of target and background to selecte features,establish feature ranking classifiers, update feature ranking classifiers adaptivly in tracking time. Kalman filter is used to predict target area, then utilize Mean-shift algorithm combined with feature ranking classifiers to complete target tracking task precisly. The tracking algorithm above can select features adaptivly according to different objectives and background informations, it is very beneficial for overcoming illumination, interference, occlusion and so on in the traffic sence. (3) Present a motion behavior recognition method based on trajectory analysis. We use the cluster method to learn movement pattern of the trajectories, get the trajectory reference sequence which represent the campaign mode. Then trajectory is viewed as a time-varying data recording the target behavior, because of dynamic time warping (DTW) technique does not limit to the length of the time series,we combinate the DTW technology and K nearest neighbor algorithm to match the trajectory which will be identified with the reference trajectory sequence of the template.In the matching process, in order to accelerate the matching speed,using DTW lower bound function to exclude all non-similar trajectory after clustering,and then matching, identifing the target moving state.Experimental results show that the object detection, object tracking algorithm can detect objiect effectively and track object stable, the behavior recognition method based on trajectory analysis achieves a higher pedestrians behavior recognition rate at the intersection,for example turning left, turning right, going forward, U-type turns.更多还原