【作者】 田彬；

【导师】 郑晓势；

【作者基本信息】 山东轻工业学院 ， 计算机应用技术， 2010， 硕士

【摘要】 智能视频监控作为计算机视觉领域的一个新兴的研究方向,在军事和生活领域拥有广阔的发展前景。智能视频监控的主要目的是对视频图像序列中的运动目标进行检测、跟踪、分类和行为分析。目标分类作为智能视频监控的一个关键步骤,其目的是对提取的运动目标进行语义上的分类,为目标跟踪或行为分析提供信息。本文主要研究基于静止单摄像机的普通户外场景下的运动目标分类。目标分类的准确性是由采集信息的质量决定的。然而,在户外场景下进行信息采集时,难免会受到光照变化、遮挡、天气等外界因素的干扰。运动目标自身姿势、位置、角度的变化也会给信息的采集带来困难。在这样复杂的外界环境下,若要做到准确的分类,同时满足智能视频监控系统实时性的特点,必须选择最能反映目标本质的特征集合,并且提高分类器的分类速度和性能。本文正是从这两个方面着手,主要进行了以下几方面的工作:①在特征选择方面,提出了一种新的构造特征集合的算法。首先,提取运动目标的多个特征,然后利用这种算法对每一种特征的分类性能进行评价,最后选出分类贡献率最高的一组特征构成特征集合。通过这种方法进行特征选择后,在保证分类精度的前提下提高了分类速度。②对局部二进制模式(LBP)进行改进,并将其用于视频目标分类。近年来,局部特征因其较强的鲁棒性,成为研究热点。局部特征在目标识别和图像匹配方面得到了广泛的应用,却很少被用于视频目标分类,主要是因为其计算量较大,且对摄像头的分辨率要求较高。LBP算子是一种有效的纹理描述子,而且计算比较简单。本文将LBP算子用于视频目标分类,取得了不错的效果。③将经典的机器学习算法AdaBoost算法引用到视频分类中,并将其用于多类分类。AdaBoost算法在构造分类器的同时还完成了对特征的选择。提取LBP算子进行分类时,特征数目庞大,使用AdaBoost算法在对LBP特征进行选择的同时完成了分类器的构造,大大提高了分类的速度和精度。更多还原

【Abstract】 As an emerging field of computer vision research direction, intelligent video surveillance is widely used in military and civil application. Intelligent video surveillance aims at detecting, tracking, identifying moving objects and understanding objects behaviors through analysis and processing image sequences. Objects classification is an important aspect of intelligent video surveillance whose research content is to classify moving objects into semantically meaningful categories and provide information for tracking or understanding objects behaviors.The moving object classification of normal outdoor scenes based on static odd-camera is studied in this paper. The accuracy of objects classification is determined by the quality of the gather information. But when the information is being collected in the outdoor, it will be affected by illumination changes、occlusion、weather and other external factors. The change of the position and angle of the objects also affect information collecting. In order to classify objects into right classes and achieve real-time in such a complex external environment, the features that most able to reflect the nature of the objects should be chose, the speed and the performance of the classifier also should be increased.Considering these two aspects, the main work of this paper is as follows:1. A new method for feature selection is proposed. Firstly, many features are extracted from the objects. Then this new method is used to evaluate the performance of each feature, and form a good sub-set of features for classification. After using this method for feature selection, the speed of the classifier is greatly improved under the premise of ensuring the classification accuracy.2. Local Binary Pattern is improved, and used for objects classification. In recent years, local features are more concerned because of its robust. Local features are widely used in objects recognition and image matching. Because of the large amount of calculation and high demanding of video resolution, it is rarely used for objects classification. LBP is an effective texture descriptor, and its calculation is very simple. LBP is used for objects classification in this paper, and the result is good.3. Adaboost algorithm, as a classical algorithm in machine learning, is used for objects classification in video surveillance. And it is used for multi-class problem. The Adaboost algorithm not only forms a strong classifier, but also selects features. When LBP is used for classification, there is large number of LBP features. The Adaboost algorithm evaluates the performance of each feature, and selects features to form a strong classifier. The speed and performance of the classification are improved.更多还原