基于轨迹关联的多目标跟踪

Multi-object tracking via trajectory association

介绍了一种基于轨迹关联的多目标跟踪算法,该算法通过两种不同的关联策略,生成跟踪目标的全局轨迹与局部轨迹,进而实现多目标跟踪。首先,基于场景自适应方法生成局部轨迹,实现检测响应与原有轨迹关联;然后,基于增量线性判决的表观模型,实现全局轨迹关联;最后,基于非线性运动模型,实现轨迹片段间空缺填补,以获取完整且平滑的跟踪轨迹。在PETS 2009/2010视频库及TUD-Stadtmitte视频库的实验表明,文中所提方法能在目标遮挡、不同目标具有相似外貌特征、运动目标方向突变等复杂情况下,实现多目标的正确关联,最终得到稳定、连续的跟踪轨迹。

A novel framework to deal with the multi-object tracking algorithm is proposed. Two different association strategies are utilized to obtain global and local trajectories by the algorithm. A scene-adaptive method is firstly utilized to generate local trajectories to construct the association between new detection responses and original trajectories. Then, a novel discriminative appearance learning method is utilized to construct the association between global trajectories. Finally, a non-linear motion model is utilized to fill the gaps between trajectories, thus obtaining continuous and smooth trajectories. Experimental results on PETS 2009/2010 and TUD-Stadtmitte database demonstrate that the proposed method can achieve stable and continuous tracking trajectories under the occlusion between objects, similar appearances on different objects and direction changes of tracking objects in complex scenes.