【作者】 刘杰；

【导师】 孙吉贵；

【作者基本信息】 吉林大学 ， 计算机软件与理论， 2007， 博士

【摘要】 本文对计算机视觉和图像智能模式识别的相关技术进行深入研究,建立了一个基于计算机视觉技术的智能乘员类型识别系统,实现了包括6岁儿童、第95百分位男性、第5百分位女性和第50百分位男性的4个类型乘员的自动识别。从而为后续拟开展的智能乘员约束系统的安全气囊控制提供有效的乘员类型信息。即智能乘员约束系统根据不同类型乘员,动态决策安全气囊的不同起爆时间和充气强度等工作模式,从而实现对不同类型乘员的最佳保护。本文主要包括对图像的预处理、图像的主特征提取方法、特征选择方法、分类器设计等方面的研究。本文的主要贡献包括如下几个方面:一、提出一种新自适应滤波算法---正交中-均值混合滤波算法,能有效去除冲激噪声和高斯噪声,同时保留大部分图像信息,与邻域平均滤波方法、中值滤波方法相比,该算法具有更好的实验效果。二、提出一种新方法ECBT(Extracting Continuity Base on Templet),用于乘员自身信息提取。该方法的思想是以车辆空载状态为比对模板,经与待处理图片进行比模处理后,达到分离背景信息,保留乘员完整信息的目的,最终以连续体的形式提取出乘员轮廓信息。三、分别采用5阶、10阶、15阶、20阶、25阶、30阶、35阶和40阶Legendre矩对乘员自身轮廓信息进行特征提取,根据对以上阶数的Legendre矩图像重构结果和计算时间,最终提取乘员图像边缘的30阶Legendre矩向量,作为乘员图像的主特征向量。四、提出一种新的混合型特征选择算法RDRK。主要包括以下几个方面:提出一种新的基于ReliefF的算法NReliefF,用于特征加权;提出一种降维算法DR,根据最小原则对加权特征进一步筛选,实现降维;采用DR算法的输出作为KNN分类器的输入,把NReliefF和KNN结合起来,采用穷举搜索策略来保证根据分类性能测量(分类正确率)选择最优特征。实验表明,采用本文提出的特征选择算法RDRK后,分类正确率和分类效率都明显提高。五、本文进行了聚类算法研究。首先对近年来聚类算法的研究现状与新进展进行了全面综述,即对近年来提出的较有代表性的聚类算法,从算法思想、关键技术和优缺点等方面进行了全面分析。在此基础之上,本文提出kNN划分的概念,进而提出基于kNN划分的聚类算法,并把传统模糊C均值算法与KNN划分相结合,形成NFCM算法。此外,本文还对标准K最近邻分类算法进行了改进,形成均值KNN分类算法,并对K的不同取值进行实验,找出了最优K值。均值KNN算法无论从分类正确率还是分类效率,都明显好于标准KNN算法。六、用多种方法实现了乘员类型识别。本文建立了包括NFCM1分类器、NFCM2分类器、标准KNN分类器、均值KNN分类器和BP神经网络分类器在内的5种分类器。我们对这5种分类器的分类正确率和效率进行比较,并对每种分类器产生的结果进行深入分析。实现了不同类型乘员的较高准确率自适应智能模式识别。更多还原

【Abstract】 As a kind of modern vehicle, automobile is entering people’s life and work. When bringing convenience and pleasure to people, automobiles threaten people’s life and safety, brings hell and gone loss to persons, by the traffic accident. Thus, the safety of the automobile has been paid more attention, and how to improve the safety performance of the automobile in the traffic accident is becoming a hot research topic for numerous researchers.The automobile occupant restraint system is composed of safety airbag and safety belt, which has to following two merits:(1) to reduce the body damage produced by the second crash between the occupant and automotive upholstery during the crash process, or (2) to avoid a second crash. However, traditional safety airbag is designed for the 95th percentile adult male in natural position, thus the deployment and inflate intension of safety airbag may produce damage to the occupants that are out of position, small females or little children. The statistic data from NHTSA suggests that every time when safety airbag save 1000 occupants, there are 57 occupants lose their precious life because of the incorrect deployment of the safety airbag and the occupant’s incorrect pose or position of sitting.To conquer the above defects of traditional airbag, the research and application of intelligent occupant restraint system is imperative. In such systems, the work modes including different deployment times and inflate intensions of the safety airbag are determined dynamically according to the real-time detection of occupant’s body feature information. And then the optimal protection for different type occupant is realized. As a result, methods for automatic occupant type recognition are very important, so are the applications in intelligent restraint systems. They are key factors to ensure occupant’s safety.Following the FMVSS208 regulation and combining the foreign occupant feature detection system with occupant type vision detection methods, we developed a new intelligent occupant type vision detection system that utilized a cheap CMOS camera. In the system, we construct the occupant measure space through the occupant edge image detection, extract and select features based on Legendre moment and a new feature selection algorithm RDRK that is proposed at this thesis. And finally the automatic occupant type recognition system based on several classifiers was implemented in industrial computer platform. Base on the system, the most valid information of different occupant’s types is provided for the safety airbag control, which is a part of the intelligent occupant restraint system.This research is supported by the National Natural Science Foundation of China under Grant No. 60473003 and 60773099; NSFC Major Research Program: Basic Theory and Core Techniques of Non Canonical Knowledge under Grant No. 60496321; National High Technology Research and Development Program 863 under Grant No. 2006AA10Z245 and 2006AA10A309.Main contributions of this thesis are the followings:1. Pre-processing of occupant image.A new self-adpative algorithm called orthogona-median filter is proposed in this paper. It is shown that the orthogona-median filter algorithm can filter the image well. The pre-process of original occupant image is realized in this phase, including both the image grey and smooth process.The orthogona-median filter algorithm proposed in this paper can wipe off the impulse noise and Gaussian noise, and at the same time most part information of the image is preserved. Compared with other filter algorithms such as neighbor median algorithm, median filter algorithm and self-adaptive algorithm, the orthogona-median filter algorithm has a better filtering result than those of others. To exploit its better performance, the orthogona-median filter algorithm is used in the smooth process of the occupant image.2. Feature extractionThe edge is the basic feature of an image .The occupant edge feature is treated as the main feature in our system. Several existing edge-detection algorithms are tried in the edge extraction process. Finally, the Canny operator is used in the occupant edge detection method, for it’s integrality of the edge linetype, the ability of noise-resist and the ability of edge orientation accurately.However, the edge image of the occupant made by Canny operator includes part of the background information. To solve this problem, a new algorithm ECBT (Extracting Continuity Base on Template) is proposed in this thesis, through which the outline information of occupant itself can be extracted out. The idea of the method is to use the picture of the empty vehicle as a contrast template. Based on the comparison between the contrast template and the image under processing, the background information is dispersed, and other information of the occupant is preserved. Finally the occupant information is extracted as continuity form.The feature vectors of occupant image are extracted using Legendre moment of 5th order, 10th order, 15thorder, 20thorder, 25th order, 30th order and 40th order, because of the characteristics for both the edge image of different type’s occupant and Legendre moment. Finally, we use the Legendre moment of 30th order to select the feature of the occupant for its better reconstruction result and time of reconstruction.3. Feature selectionA new feature selection algorithm named RDRK is proposed. The RDRK algorithm is shown to be better than ReliefF algorithm. The main task of this phase include: (1) feature pre-selection: The amount of the feature extracted by the Legendre moment is very large, so we adopt the NReliefF algorithm to evaluate a weight on each feature, and then the features are filtered by the algorithm DR based on minimized rule; (2) feature selection: The result of the feature pre-selection is taken as the input of the KNN classifier, and then delivered to a combination of NRelief and KNN. Exhaustive search strategy is used to select the optimal features based on the highest classify accuracy. Experimental results show that the adoption of the RDRK algorithm brings a highly improvement on both classify accuracy and efficiency. Specifically, the feature selection process consists of the following steps:①Each data includes 961 dimension features is processed by a feature extraction method that uses 30th order Legendre.②In the initial phase, the same feature in each feature vector is cut off, and the number of remain feature is 469.③The NReliefF algorithm is used to give weight for 469 features each, and the number of feature vectors is KK*R.④Then the feature dimension reduction is finished by DR algorithm based on minimized rule, i.e., if there is one negative value of the feature appeared in one of the KK*R vectors, then this feature is cut. So the result of DR is 197 dimension feature preserving.⑤Input these 197 dimension features to the KNN classifier, keep the most important 47 features which lead to the highest classify accuracy.4. Research on clustering algorithmsFirstly, we give a through review on the current progress of clustering algorithms developed in recent years. Several representative clustering algorithms are introduced and analyzed from different aspects, including the main ideas, the key technology, advantages and shortcomings.Secondly, the concept of KNN partition and a clustering algorithm based on KNN partition are proposed. The later is then combined with the FCM algorithm to bring a new classification algorithm called NFCM. In addition, based on the standard KNN algorithm, we developed a mean KNN classification algorithm (NKNN). The optimal K value is found by experimentations. The mean KNN classification method has a better performance in both classification accuracy and efficiency than that of the standard KNN algorithm.5. Classifier designTotally, 5 classifiers are developed in the paper: the NFCM classifier, NKNN classifier, FCM classifier, KNN classifier and BPNN classifier. Both the accuracy and efficiency of these classifiers are evaluated, and each result is deeply analyzed. Based on the best classifier NKNN, we develop a self-adaptive pattern recognition system, whose recognition accuracy reaches 93.4343%.6.Future work●More images of different types will be collected to enlarge the size of training set.●Based on computer vision information, we will exploit the information of the weight sensor to improve recognition rate of occupant type.更多还原