复杂场景中数目变化的视觉多目标实时跟踪技术研究

项目名称： 复杂场景中数目变化的视觉多目标实时跟踪技术研究

项目编号： No.60802067

项目类型： 青年科学基金项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 刘宏

作者单位： 中国科学院计算技术研究所

项目金额： 18万元

中文摘要： 视觉多目标跟踪是视觉运动分析关键技术之一，广泛应用于视频监控、视频编码、运动合成等领域。然而复杂背景和光照变化的干扰，以及数目变化的目标运动所产生的目标分块、合并和分离等问题，使跟踪系统很难在实时性、稳健性以及准确性上达到平衡。针对复杂场景下数目变化的多目标跟踪所面临的难点问题，本项目主要研究了以下内容：1.针对复杂场景的运动前景检测和分割技术研究；2.基于直线特征的行人和车辆分类研究；3.融合静态和动态特征的快速鲁棒的行人检测技术研究；4.融合Meanshift和粒子滤波的多目标跟踪技术研究；5.基于运动检测和粒子滤波的变化数目的多目标跟踪算法。通过以上核心算法研究，提出了一系列新颖的方法，在若干公开测试集上验证了算法的有效性。基于本项目的研究建立了一个鲁棒的视觉多目标跟踪系统，能对复杂场景中数目变化的运动目标进行有效的检测和跟踪，为目标行为识别和事件检测提供信息和决策支持。在研究成果的推广和应用方面，将本项目的研究成果应用于一个实际的监控场景，进行违章事件的检测，得到相关部门的认可。本项目的研究对于推动视频运动分析的发展和实用化具有重要的理论意义和实际价值。

中文关键词： 多目标跟踪;目标建模;运动目标分割

英文摘要： Multi target tracking is a key technology of visual motion analysis. As one core problem in field of computer vision, visual target tracking has wide application in video surveillance, video coding, sports video analysis, etc. However complex background and illumination change and especially vary number target will result in merger and separation of the target blocks, which make it difficult to achieve the balance on the accuracy and robustness of whole tracking system. This project aims to analysis and solve some difficult problems faced by vary number target tracking in complex scenes. The main research contents of this project include: 1. Motion foreground detection and segmentation technology in complex scenes; 2. Pedestrians and vehicles classification research based on Hough line features; 3. The rapid robust of pedestrians detection fusion static and dynamic features; 4. Multi target tracking based on Meanshift and particle filter; 5. Vary number target tracking based on motion detection and target tracking algorithm. This project proposes a series of novel methods for above research. Experimental results and performance evaluation on several open test dataset verify the effectiveness of the proposed algorithms. Based on research of this project, we build up a visual target tracking system, which can perform vary number target tracking in complex scene. These results can provide information and decision support for behavior identifying and incident detection. As one application of this research, we have adopted this system in a real station visual surveillance scene to detect incident issues. The research of this project will help to promote the development of visual motion analysis.

英文关键词： multi target tracking; target model; motion detection and segmentation