Abstract: As one of the pillar techniques in intelligent transportation system, vehicle tracking has attracted wide attention of many scholars. SIFT (scale-invariant feature transform) features can effectively solve problems such as rotation, scaling, translation of target and provide firm support for tracking vehicle. But traditional SIFT features are unable to distinguish foreground from background, centralism of too many matching features on the background will result in the loss of tracking target. Based on tracking SIFT features, this paper introduces the vehicle tracking method based on SIFT features and GrabCut algorithm. SIFT features can effectively solve problems such as posture change and variance of vehicles in the near to far field, GrabCut algorithm can distinguish foreground from background. The experimental results show that if the camera is not shaking obviously in the day, the initial frame motion detection can realize stable tracking for the moving vehicle and effectively solve problems such as posture change and variance of vehicles in the near to far field.

Key words: vehicle tracking, scale-invariant feature transform (SIFT), GrabCut

摘要： 车辆跟踪作为智能交通系统中的一项关键技术备受广大学者关注。SIFT（scale-invariant feature transform）特征可以有效解决目标的旋转、缩放、平移，为车辆跟踪提供很好的特征支持，但是传统的SIFT特征跟踪不能区分前景和背景，极多的匹配特征集中在背景上，导致跟踪目标丢失。在研究现有车辆跟踪算法的基础上，提出了基于SIFT特征与GrabCut算法的车辆跟踪方法，SIFT特征有效解决了车辆姿态变化及远近变化问题，GrabCut算法有效保证了前景及背景的准确分割。实验表明，该方法在日间摄像机不明显晃动环境下，初始帧运动检测车辆后能够对运动车辆实现稳定的跟踪，并且有效解决了车辆姿态变化及远近变化问题。

关键词: 车辆跟踪, 尺度不变特征转换（SIFT）, GrabCut