基于数据与模型混合驱动的密集人群中特定人脸持续跟踪方法

As the widely usage of video survilliance system, research for long-term tracking of faces in crowds people in complex secen (i.e. quickly filter suspect faces and persistent tracking in complex scene) is becoming a key point in video tracking. Deterministic hypothesis tracking based on data-driven or model-driven is not adequate in complex scene and target motion uncertainty. In this project, a hybird dirven tracking method using data-based and model-based concepts with online close-loop learning scheme is proposed. Combing with chaotic features extracted from videos, probability confidence detection and probability hypothesis density mehtods. The critical point of this research is to deeply explore persistent tracking for specific traget in crowds people: obtaining offline prior information of chaotic features for specific target; probability confidence detection for specific face detection criterion in strong disturbance; a persistent single object tracking algorithm based on SIFT tracking with online learning and detection. Firstly suspects are detected in crowds people. Then feature updating and peresistent tracking are achived by online learning and SIFT-based tracking. This work provides a basic support for persistent tracking in complex scene and is a preliminary exploring to persitent estimate target in unstructured enviroment.

随着视频系统应用需求的深入，研究复杂场景内密集人群中特定人脸持续跟踪方法（即：如何在复杂场景中快速筛选出与数据库中相似的疑似人脸并对其持续跟踪）成为视频跟踪的关键问题之一。由于复杂环境和目标运动的不确定性，单纯基于数据驱动或模型驱动的确定性假设条件的跟踪算法时常显的不能适应，效果不佳。本项目采用数据与模型混合驱动的在线学习闭环跟踪框架，结合视频混沌特征分类、概率置信度检测和概率假设密度方法，重点研究复杂场景下的密集人群中特定目标持续跟踪方法的深层次问题：基于混沌分类的目标离线先验信息获取方法；强近似干扰下概率置信度特定目标人脸检测准则；基于SIFT跟踪环的在线学习-检测单目标持续跟踪算法。通过在密集人群中检出疑似目标，利用在线交互学习和SIFT跟踪环实现目标特征模型更新和持续跟踪。为解决复杂场景下特定目标持续跟踪提供应用基础支持，同时也对非结构化环境系统中目标的持续估计方法进行初步探索。

由于复杂感知环境、目标区域运动和目标数目变化等因素制约，单纯基于数据驱动或模型驱动的确定性假设条件的跟踪算法时常显的不能适应，效果不佳。本项目采用数据与模型混合驱动的在线学习闭环跟踪框架，结合视频混沌特征分类、概率置信度检测和概率假设密度方法，重点研究复杂场景下的密集人群中特定目标持续跟踪方法的深层次问题。改进工作主要从补充跟踪算法航迹识别功能，提高观测似然函数的区分性和可靠性，以及提高算法状态抽取的可靠性和准确性三个方面着手对基于随机有限集的视频目标跟踪算法进行研究。基于混沌分类的特定目标离线先验信息获取方法；强近似干扰条件下概率置信度特定目标人脸检测准则；基于SIFT跟踪环的在线学习-检测单目标持续跟踪算法。通过在密集人群中检出疑似目标，利用在线交互学习和SIFT跟踪环实现目标特征模型更新和持续跟踪。实现了复杂场景下密集人群中特定目标的持续跟踪。利用城市监控真实数据，对密集人群中特定目标跟踪算法进行验证，说明了算法的有效性。