基于行为模型和超图匹配的多目标跟踪技术研究

Due to the lack of correct modeling of object kinematic characteristics and interaction between the objects and their environment, current multi-target tracking (MTT) algorithms often fail in dense or highly dynamic scenarios, where nearby similar objects frequently lead to tracking errors or loss of tracks. In order to more properly represent object motion for better trajectory prediction, we propose a novel MTT approach based on a social behavior model and hypergraph matching. First, we leverages more sophisticated social behavior models to capture kinematic characteristics of the objects and the naturally formed groups under the environmental constraints. This provides more accurate motion prediction that can well approximate the real world situations and significantly reduce association searching range in the following steps. Second, a relation graph is employed to model the relationship among objects and the constraints from surrounding environment, which incorporates both vertex and soft edge matching schemes into single cost minimization framework to obtain better object association between consecutive frames. Last, using a sampling method named Markov Chain Monte Carlo Data Association, we dynamically refine the detected trajectories locally within a sliding time window, so that the temporal constraint is emphasized and the number of track breaks due to occlusion is reduced. This additional step will further help our system efficiently converge to an energy optimum.

现有的多目标跟踪算法由于没有充分考虑不同目标类型自身所特有的运动特性、目标间的互动关系、以及环境对目标运动约束的影响等，在运动目标密集的情况下很容易受到邻近的相似目标的干扰而造成目标丢失或跟踪错误。为了能够更为准确地描述目标运动状态和预测目标未来的运动轨迹，本项目提出基于行为模型和超图匹配的多目标跟踪算法。该算法首先对目标、目标群、以及目标运动环境下的群体行为进行分析建模，使目标的运动更为贴近运动物体的真实物理属性；随后，我们利用二元关系图匹配模型建立对目标间及环境对目标约束间相互作用关系的描述，强化多目标运动的时空约束关系，更好地实现帧间多目标关联；最后，借助基于马尔可夫链蒙特卡洛方法的数据关联算法和滑动窗口技术，我们对所生成多目标轨迹进行的调整和局部关联，强化运动轨迹在时域上约束关系，减少由于目标被连续遮挡而造成的轨迹断裂，实现多目标跟踪轨迹的全局最优逼近。

现有的多目标跟踪算法由于没有充分考虑不同目标类型自身所特有的运动特性、目标间的互动关系、以及环境对目标运动约束的影响等，在运动目标密集的情况下很容易受到邻近的相似目标的干扰而造成目标丢失或跟踪错误。为了能够更为准确地描述目标运动状态和预测目标未来的运动轨迹，本项目提出基于行为模型和关系图匹配的多目标跟踪算法。该算法首先引入微观运动模型对目标及其环境下的行为进行分析建模，使目标的运动更为贴近运动物体的真实物理属性；随后，我们利用二元关系图匹配模型建立对目标间及环境对目标约束间相互作用关系的描述，强化多目标运动的时空约束关系，更好地实现帧间多目标关联；最后，通过对轨迹片段的图分析工作，我们提出了平行和串行轨迹片段图来描述各个轨迹片段间的相互关系和作用影响，保证了在时空域上发现各个轨迹间组群关系和轨迹片段的完整性和连续性。实验结果证明该算法能有效地提高多目标跟踪和轨迹关联的精度，能够为后续的目标行为分析等应用提供支撑。