目标检测是计算机视觉的重要研究内容。本文针对基于边缘轮廓信息统计学习的目标检测方法展开讨论。在边缘轮廓特征提取和判别式分类的目标检测方法基础上，本文针对特征共享方法、可形变目标精确检测方法、人脸检测具体方法以及眼睛轮廓检测跟踪具体方法四个方面进行深入研究。本文主要贡献如下：(一) 提出边缘特征和共享Boost分类器并将其应用于多任务目标检测。首先提出一种积分边缘梯度特征以提取目标的边缘轮廓信息。其次基于目标检测多分类器遍历搜索的特点，提出特征可共享的数学条件；并在此基础上基于Boost框架和遗传规划提出共享Boost分类算法。该方法可利用不同类别/子类之间的相似区域来减少特征数并提升分类性能。最后提出融合多部件共享Boost输出的鉴别式部件模型用于目标检测。该模型可提高低分辨率和自遮挡条件下的目标检测性能。(二) 提出基于轮廓特征的可形变目标的精确检测方法。针对轮廓特征提取的问题，使用首尾相接的线段拟合轮廓片段并使用边缘图上的矩形近似计算轮廓的投影梯度强度。为精确检测目标位置，基于主成分分析获取目标边界的统计形状模型并定义形状参数。最后提出分类-回归Boost算法来同时获取目标/非目标区域的判别参数和目标形状参数。本文方法可实现可形变目标的边界精确定位检测。(三) 提出融合头肩轮廓信息的人脸检测方法，并提出基于相似半脸模式的多姿态人脸检测方法及基于随机森林回归融合多特征的姿态估计方法。针对复杂条件下人脸检测性能下降的问题，提出融合头肩轮廓信息来补偿人脸模式缺失。该算法可大幅提高复杂光照下的人脸检测精度。针对姿态估计问题，首先提出利用相似半脸模式构建树状结构分类器的多姿态人脸检测方法；然后利用随机森林回归融合多特征获取姿态角估计值。本文方法可获取精确连续的人脸姿态估计结果。 (四) 提出基于边缘的眼睛轮廓检测与跟踪方法。首先提出改进抛物线Hough变换以获取精确的眼睛轮廓位置。同时为在视频中检测眼睛状态，提出基于粒子滤波跟踪框架和边缘特征的眼睛轮廓跟踪方法。该方法根据眼睛运动和状态转换特点构建动态模型，并结合边缘特征构建观测模型。本文提出的眼睛轮廓检测与跟踪算法可以获取精确的眼睛轮廓检测并提升眼睛状态判别的准确度。

Object detection is an important problem in computer vision. In this dissertation we concentrate on the research of object detection methods based on statistical learning of edge information. Based on the extraction of edge/contour features and classification algorithm, we mainly focus on four issues as feature sharing methods for multi-task object detection, deformable object detection methods with contours, face detection and its applications, eye contour detection and tracking. The main contributions are as follows.Firstly, we propose shared-Boost algorithm for multi-task object detection. A novel kind of descriptors called integral oriented gradient features is proposed to extract the edges/contours of objects. Besides, a mathematical definition of shared features for sliding-window based object detection methods is given. Based the definition, shared-Boost algorithm is proposed. Similar regions of multiple object classes can be extracted with shared-Boost. To apply it into part-based object detection, we propose a discriminative model which fuses the outputs of shared-Boost into the confidence of objects. The proposed shared-Boost based discriminative part models show efficiency, especially in low-resolution and partially inter-human occluded objects. Second, we propose Classification-Regression Boost and construct a framework for deformable object detection with contour features. We propose a novel kind of contour descriptors called fitted contour fragment features. Several line fragments are used to fit the contour fragments and sums of edge intensities in rectangle areas are used to approximately compute the response of contours. To locate object precisely, several points along object boundaries are used to represent object locations and principle component analysis is applied to construct statistical shape models of objects. The object boundaries can be recovered with boundary parameters and trained shape models. A novel kind of boosting named as classification and regression boosting to discriminate object/non-object regions and recover boundary parameters at the same time. The proposed methods can be applied in detecting deformable objects.Third, we propose a method to fuse head-shoulder contour information for face detection under complicated conditions. The performance of face detection decreases a lot under complicated conditions such as bad illumination and occlusion. To deal with the problem, a head-shoulder detector is trained with integral oriented gradient features. A probabilistic model to fuse the output of face detector and head-shoulder detector is presented. By adding contour information, the performance increases a lot. Also we propose a framework which integrates tree-structured multi-view face detection and head pose estimation. Tree-structured cascaded-Adaboost classifiers which preserve similar patterns of faces under various poses are trained for multi-view face detection. Random forest regression with multiple features is applied to estimate pose angles precisely. Finally, a framework to extract eye boundaries with edge features is proposed. It integrates both eye contour detection and tracking. Two parabolas are used to fit eye boundaries. A modified Hough transform for parabolic curves is proposed to extract upper and lower eyelids. We use the coefficient of parabola as the variant and the locations of vertex change along with it. Besides, a framework to track eye/eyes contours based on particle filtering and edge features is also proposed. We construct a dynamic model for eye movements and eye state transition. Edge features are applied to construct the observation model. With proposed eye contour detection and tracking algorithm, the accuracy of eye state detection can be increased.