【作者】 李子龙；

【导师】 刘伟铭；

【作者基本信息】 华南理工大学 ， 交通信息工程及控制， 2014， 博士

【摘要】 视频目标的检测、识别是目前智能交通和计算机视觉领域中的一个重要研究方向。但是，由于检测和识别环境下存在背景复杂、光照变化、目标遮挡等原因，导致该应用仍面临着许多困难，检测和识别的鲁棒性及准确性都有待进一步提高。本论文对视频目标检测和识别中的几个关键问题进行了研究，主要包括：复杂场景下目标与背景、阴影的准确分割；对提取的前景目标准确分类；复杂背景下的目标识别。针对这些问题，本论文提出了相应的解决方法。具体工作如下：1.提出了一种基于自适应模糊估计的背景建模方法。该方法从函数估计的角度对背景进行建模，并采用TSK模糊系统作为估计函数。为了训练函数估计算子，分别使用粒子群优化（PSO）算法和递归最小二乘估计（RLSE）算法来优化模糊系统的前件参数和后件参数。为了有效估计背景，将前景像素看作背景像素的异常样例，并提出了异常样例的去除方法，然后用去除后的结果去训练模糊估计算子。该方法在动态背景、光照变化、摄像机振动等环境下都具有较高的运行效率和检测效果。2.提出了一种基于模糊积分的运动阴影检测方法。在提取前景区域的基础上，选择颜色和纹理作为阴影检测的特征，并分别定义了这两种特征的相似性和重要性测度函数，然后通过Choquet模糊积分将这两种特征融合，实现阴影和前景目标的分类，最后通过后续处理，找到真正的阴影区域。3.提出了一种基于JointBoost I2C距离度量的目标分类方法。针对经典I2C距离计算量大且易受噪声干扰等不足，首先提出了一种原型特征集的生成方法，该集合中的样本数量较少，但更具有代表性，计算测试图像到该原型特征集的距离花费较少时间；然后借助JointBoost算法的思想，联合多个I2C距离度量生成一个强分类器；最后还提出了一种将空间信息融合到强分类器的方法。实验证明，该方法在前景目标和图像分类实验中，具有更高的分类性能。4.提出了基于特征码本树和能量最小化的目标识别方法。该方法考虑了特征的空间位置信息和特征之间的空间关系，集成了目标检测和目标识别。首先从目标图像提取的大量特征中过滤掉噪声特征；然后对单特征和空间上邻近的串联双特征分别使用层次k均值聚类算法构建特征码本树，利用树模型可以实现特征快速定位和分类；最后建立一个能量函数来融合单、双特征码本树的类别概率匹配结果，并通过在测试图像中寻找滑动窗口所在区域的能量最小化来确定所属类别目标的位置。5.提出了基于优化Hough森林代价损失的目标识别方法。首先在充分利用训练图像中对象位置是已知的基础上,提出了改进的偏移量不确定性度量方法；其次借助Boosting算法的思想，学习图片块样本和目标对象样本的自适应权重分布，并分别优化用于构造随机树和Hough森林的代价损失函数；最后根据图片块样本的权重分布，提出了改进的类标志不确定性度量方法。基于Hough森林的代价损失函数，还提出了随机树权重的学习方法。更多还原

【Abstract】 Visual object detection and recognition is an important research direction of theintelligent transportation and computer vision.However, complex background, illuminationchanges, object shielding and so on in practical detection and recognition environment is stillfacing many difficulties, so the robustness and accuracy of handling these proplems should befurther improved.Some key issues of visual object and recognition are researched in this dissertation,which mainly include: accurate segmentation of background, objects and shadows undercomplex scene, accurate classification of the extracted foreground objects, object recognitionunder complex background.Motivated by these isssues, the corresponding solutions areproposed in this thesis.Specific studies as follows:1. A new adaptive fuzzy estimate method for background modeling is proposed.Thismethod models background from the perspective of function estimation, and uses TSK fuzzysystem as the estimated function.In order to train the function estimator, the proposedapproach combines both the particle swarm optimization (PSO) and the recursive leastsquares estimator (RLSE) to optimze the parameters of the premise part and consequent partof the fuzzy system. In order to effectively estimate background, we interpret foregroundsamples as outliers relative to the background ones and so propose an outlier Separatormethod. After the outliers are removed, the obtained results are used to train the fuzzyestimator. This method has the high accuracy and effectiveness in dynamic background,illumination changes, camera vibration and so on.2. A new shadow detection method based on choquet fuzzy integral is proposed. Afterthe foreground area is ectracted, this method chooses color and texture as detection features,and defines their similarity and feature importance degree measure functions. Then, thesemeasures are integrated by using the choquet integral in order to initially find shadow andforeground. Finally, the real the shadow areas is distinguished after the subsequent processingstep.3. A new visual object classification method based on JointBoost I2C distance metric.Because the I2C distance has the expensive computation cost, and it is also easily affected bynoise features. At first, we propose to generate a prototype feature set, which has fewersamples, but more representative.The distance calculation of test image to this set can spendless time. And then, we adopt the idea of JointBoost algorithm, combine multiple I2C distance metric to generate a strong classifier, and also propose a spatial information fusion method.Experimental results show that the proposed method has higher classification performance inforeground object and image classification.4. A new visual object recognition method based on feature vocabulary tree and energyminimization is proposed. This method takes into account the spatial position information oflocal features and the spatial relationships between local features, and combines objectdetection and classification. At first, a large number of features extracted from object imageare filtered. And then, the single feature and concatenated pairwise feaures are respectivelyused to build the vocabulary trees by using hierarchical k-means clustering algorithm, thetree-structure has the advantages of the fast lookup feature localization and calssification.wefinally built an energy function combing the matched class probability results of these twovocabulary trees, and then the visual object location and object type are known by minimizingthe energy function of the sliding window in test image.5. A new visual object recognition method based on optimizing Hough forest cost lossfunction is proposed. At first, we propose an improved offset uncertainty measure, exploitingthe knowledge of the object locations in the training images. And then, we adopt the idea ofBoosting algorithm to learn the adaptive weight distribution over the image patch samples andobject image samples, and respectively optimize the cost loss function of constructing randomtree and Hough forest. At last, we propose an improved class uncertainty measure by usingthe weight distribution of the image patch samples, and also give a learning method ofrandom tree weight based on Hough forest cost loss function.更多还原