交通场景下基于深度和迁移学习的行人检测与跟踪方法的研究

项目名称： 交通场景下基于深度和迁移学习的行人检测与跟踪方法的研究

项目编号： No.61471154

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 孙锐

作者单位： 合肥工业大学

项目金额： 81万元

中文摘要： 车载行人检测系统是一个运动的检测平台，交通场景的多样性和时变性给行人检测与跟踪带来一系列的技术难题。本项目旨在创新行人目标的分层深度表示模型，探索变化场景中的行人检测与跟踪方法、对由此引出的关键科学问题展开深入系统的研究。针对交通场景下行人尺度变化较大、频繁遮挡的情况，基于深度学习实现一种新型的分层视觉特征学习方法与计算模型；针对在源场景下训练得到的分类器在变化场景下性能快速下降的问题，提出从源场景和目标场景两个角度建立样本的迁移模型，探索基于流形学习和多线索的可信样本选择机制，研究集成可信得分的加性核SVM分类器设计方法。面向场景变化下的目标跟踪漂移的问题，将稀疏表示方法引入粒子滤波器框架，创建一种新的具有场景变化适应能力的目标跟踪计算模型和目标模板的在线更新策略，鲁棒地适应目标外观变化与遮挡的影响。研究的内容和方法为汽车主动安全与无人驾驶提供了核心技术支撑，具有重要理论意义与实用价值。

中文关键词： 图像识别；行人检测；目标跟踪；深度学习；迁移学习

英文摘要： Vehicle mounted pedestrian detection system is a dynamic detection platform. The diversity and time-varying of traffic scenes make the pedestrian detection face to a series of technical difficulties. The project aims to propose hierarchical deep representation model, explores pedestrian detection and tracking techniques in changing scenes, and researches in-depth the key scientific problems behind the techniques. Considering the pedestrian scale changing and frequency occlusion, the project implements a novel hierarchical visual features learning method and computation model based on deep learning. Because the classifier which trained in the source scene fast decreases the detection performance in changing scene, the project proposes the samples transfer models from source scene and target scene and explores the confidence samples selecting mechanism based on manifold learning and a set of context cues. The confidence scores are formulated design method of additive kernel support vector machine (SVM)classifier. Considering the target tracking drift problem, the project introduces sparse representation into particle filtering framework, proposes new target tracking method robust to scene change. The method can adapt robustly to the effect of appearance change and occlusion. The content and methods of research provide key technical support for vehicle active safety and autonomous driving, have important theoretical significance and practical value.

英文关键词： image recognition;pedestrian detection;object tracking;deep learning;transfer learning