基于弱监督学习和深度信息的目标跟踪算法研究

项目名称： 基于弱监督学习和深度信息的目标跟踪算法研究

项目编号： No.61202299

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 计算机科学学科

项目作者： 钟必能

作者单位： 华侨大学

项目金额： 23万元

中文摘要： 目标跟踪以其重要的理论和实用价值一直得到广泛关注。而复杂场景中遮挡、背景干扰和目标表观的变化等因素会给目标跟踪带来极大挑战。本课题针对这些问题，将快速目标检测技术和目标的深度信息融入到目标跟踪过程中，以实现基于多线索融合的鲁棒目标跟踪。具体地，本课题拟开展以下三个方面的工作：1）基于广义霍夫变换和偏最小二乘法，提出一个快速、鲁棒的目标检测算法；同时进行在线学习，以达到在目标跟踪过程中不断自适应地学习目标的表观模型、提高目标检测的正确率，并通过在线的目标检测从而不断修正目标跟踪的结果；2）开发利用目标的三维深度信息，充分挖掘深度信息在目标跟踪中判定和处理遮挡、区分前景和背景等方面的作用；3）在弱监督学习框架下，融合深度信息、目标检测、颜色和形状等互补线索，发挥各个线索的优势，实现鲁棒有效的目标跟踪。本研究对视频监控、人机交互等领域有较高的学术与应用价值。

中文关键词： 目标跟踪；目标检测；三维深度信息；多线索融合；弱监督学习

英文摘要： Object tracking has been received extensive attention, due to its important theoretical and practical value. However, the challenges in complex scenes, such as occlusion, background clutter and object appearance changes, still bring great difficulties to object tracking. To address these problems, the project combines rapid object detection and object depth information into the object tracking process. Specifically, the project will focus on the work in the following aspects: 1) to propose a rapid and robust object detection algorithm based on the generalized hough transform and partial least squares. Meanwhile, to improve the object detection rate by online learning object appearance model; 2) to exploit the three-dimensional depth information of objects which will be used to improve the ability of the tracking algorithm in handling occlusion, distinguishing foreground objects from background scene, etc. 3) to consider object tracking in a novel weakly supervised learning framework, in which multiple complementary cues such as depth information, object detection, color and shape, are fused to achieve robust and effective object tracking. The project has high academic and applicable value in video surveillance and human-computer interaction.

英文关键词： object tracking；object detection；3D depth information；multiple cues fusion；weakly supervised learning