基于Grassmann流形的粒子滤波多目标跟踪方法研究

In complex background, building accurate and reliable appearance feature model is the key to achieving stable tracking result for deformation object, while existing flat European vector space has been unable to describe the appearance feature of deformation object area under different posture conditions accurately. Grassmann manifold not only has the space expression of smooth curved surface, but also has appropriate metrics for deformable object. The project researches the method for building the appearance feature of deformation object based on Grassmann manifold, taking use of the intrinsic geometry characteristics of state space to design the particle filtering algorithm for predicting appearance feature, it also researches affine transformation manifold to establish the geometric deformation model, and design the particle filtering algorithm for estimating the object geometric deformation. Combining the two filters, alternately tracking process and online study process, it realizes object tracking. For the situation of object obscured by the background, the project researches on the processing strategy against occlusion based on Grassmann manifold. In the process of online study for object appearance feature space, abnormal information will be shielded, to guarantee the accuracy of the feature space. For mutual occlusion problem in multi-object tracking process, it researches the bilateral filtering based on Lie group, and establishes methods for corner detection and classification, designing the algorithm for separating the objects correctly in occlusion area, even can distinguish the objects changing its motion direction after occlusion to accomplish multi object tracking stably.Finally it researches the method for adaptive particle filtering on manifold to improve computational efficiency.

复杂背景下建立准确可靠的表观特征模型是实现形变目标稳定跟踪的关键。现有的平直欧式向量空间已无法准确表达不同姿态下的形变目标表观特征。Grassmann流形不仅具有光滑曲面的空间表达方式，且具有更为适合形变目标表观特征的度量。本项目基于Grassmann流形，研究形变目标表观建模方法，利用状态空间的内蕴几何特性，设计估计表观特征的粒子滤波算法。研究仿射变换流形，建立目标的几何形变模型，设计预测几何形变的粒子滤波算法。结合两类滤波器，交替进行跟踪过程与在线学习过程，实现目标跟踪。研究基于Grassmann流形的背景遮挡处理策略，在目标表观特征空间在线学习过程中，屏蔽异常信息，保证特征空间的准确性。针对多目标互遮挡情况，研究流形上的角点检测与分类算法，正确区分目标重叠区域内各目标，特别是遮挡分裂后运动方向突然发生改变的目标，实现多目标稳定跟踪。最后，研究流形上的自适应粒子滤波算法，提高计算效率。

针对复杂背景下，跟踪经历大幅度位姿变化的目标易发生偏移的难题，本项目将Grasmann流形引入视频目标跟踪领域，重点对几何形变目标的表观建模、目标经历遮挡时的处理策略、多目标互遮挡时的分离跟踪算法、自适应的粒子滤波方法等开展了卓有成效的研究与攻关。.研究了经历明显几何形变目标的跟踪 。首先，研究了基于Grassmann距离度量的改进的谱聚类分析算法，在流形空间上分析数据点之间的相似性。然后，设计了基于Grassmann流形的动态目标的表观模型，作为粒子滤波中的观测模型，应用投影变换群描述目标的几何形变，设计粒子滤波算法完成目标跟踪。. 研究了目标经历遮挡时的在线流形学习方法。针对目标在跟踪过程中经常发生的目标被遮挡时，造成目标特征空间模板漂移，导致跟踪结果不准确或是失效的情况，研究了几何形变目标在经历遮挡时，基于Grassmann流形的在线流形学习方法，提出特征空间更新策略。能够正确判断出是否发生遮挡确保特征空间的准确性。. 针对多目标的互遮挡问题，研究了多目标互遮挡时的处理策略。首先，基于深度学习理论，研究了目标基于深度谱卷积神经元网络的高效视觉目标跟踪算法。然后基于李群流形，设计正确区分互相遮挡的各个目标的算法。并将其与所设计的跟踪算法融合，进行仿真实验，验证多目标互遮挡情况下跟踪的有效性。. 研究了自适应的粒子滤波算法。提出了基于Grassmann流形的样本数自适应调整的粒子滤波算法。在跟踪过程中，自适应的改变粒子数目，提高计算效率。. 项目的研究取得了比较丰富的研究成果，各方面研究进展均完成或超过了预定目标指标，为目标跟踪领域提供新的研究思路和关键技术，研究成果已经得到专家的肯定并得以推广运用。