【作者】 王强；

【导师】 朱梦宇；

【作者基本信息】 北京理工大学 ， 生物医学工程， 2015， 硕士

【摘要】 目标跟踪技术在人机交互、智能交通、视频监控等领域发挥着重要的作用,但由于场景的复杂性,还需要对该技术进行深入的研究。目标的姿态变化、外界光照变化、遮挡、高速运动等都会给目标跟踪带来干扰,因此构建一个稳健的跟踪系统是我们研究的方向。本文详细介绍了以TLD(Tracking-Learning-Detection)为框架的目标跟踪方法。TLD跟踪算法由检测模块、学习模块、跟踪模块三个部分组成,对于长时间跟踪具有很强的适应性。本文选取三个复杂度不同的场景对TLD跟踪算法进行仿真,实验结果表明跟踪算法在目标丢失后能够重新找回,具有一定的鲁棒性。然而,在遇到遮挡、光照变化和姿态变化时,TLD跟踪算法存在对光照和姿态变化敏感,遮挡后恢复捕捉时间长等问题。为此本文提出了一种基于Mean Shift的TLD融合算法,根据TLD跟踪框置信度来动态设置Mean Shift算法的迭代起始点,最终准确的获得目标位置。同样的,对改进算法在三种场景下进行了仿真。从实验结果来看,改进的TLD跟踪算法在目标姿态变化和外界光照变化的情况下,跟踪状态良好。当目标丢失后重新出现在视频区域时,改进算法能快速的跟踪到目标。从评估结果来看,目标框的偏移误差大大减小,重叠度大大提高。所以,改进的TLD跟踪算法能够有效的适应复杂场景的目标跟踪,较好的解决了跟踪过程中的漂移和丢失现象,使得跟踪系统更具有鲁棒性、精确性和实时性。更多还原

【Abstract】 Target tracking plays an important role in areas such as human-computer interaction, intelligent transportation, video monitoring, but we should do some in-depth research of the technology because of the complicated scene. It will bring disturbance because of the target’s posture variation, illumination change, object occlusion and high speed movement, so building a robust tracking system is our research direction.This paper mainly studies the target tracking algorithm with the system framework of TLD(Tracking Learning Detection). TLD algorithm could achieve long-term online target tracking, which is mainly composed of three parts: the tracker, learning module, the detector. This paper selects three scene with different complexity to simulate the TLD tracking algorithms, results show that the track can be rediscovered after lose, and the algorithms has a certain robustness. However, TLD tracking algorithms is sensitive to light and posture variation, long time to rediscover tracker after missed and so on when in occlusion, occlusion or posture variation scene. A TLD model and Mean Shift based tracking algorithms, a refine tracking object can be got by dynamically setting the retrieved initial point of Mean Shift with the degree of TLD confidence. Similarly, the new tracking algorithm is simulated in three scenes.The experimental results show that the improved TLD target tracking algorithm can track the target effectively when posture variation and illumination change. The target will be captured quickly when reappear. The localization error of bounding box is greatly decreased, and the overlap degree is greatly increased. So the improved TLD target tracking algorithm can effectively adapt to the complexity of the environment object tracking, at the same time be able to the drift of the target and the loss of the target,it makes the tracking system more robust, accurate and real-time.更多还原