Research On Object Detection And Classification In Intelligent Transportation

With the improvement of technology as well as the decline in the cost of hard-ware, computer vision and pattern recognition techniques have been greatly applied into transportation. Object detection and object classification in intelligent trans-portation are studied in this dissertation. Based on specific applications (i.e. human detection, car detection and traffic sign classification), the challenges are analyzed and the corresponding solutions are proposed. The proposed and related theories and methods in this paper also contributes to other computer vision and pattern recogni-tion problems. Firstly, feature extraction methods in object detection are studied and two local binary pattern based features are proposed to describe objects. Secondly, in the field of object classification, one of ensemble classifiers Boosting has achieved great successes since the first real-time face detector with AdaBoost proposed by Viola and Jones. However, conventional Boosting algorithms need a large number of training samples to train a reasonable classifier. The performance will drop dramatically if the training samples are scare. In this dissertation, many different kinds of data (e.g. unlabeled data) are combined to train better Boosting algorithms to deal with small sample size problem.The highlights and main contributions of this dissertation include:1) In the field of object detectionObject (e.g. human, car) detection is always an important component of Intel-ligent Transportation System (ITS) and Intelligent Surveillance System (ISS). Com-puter vision and pattern recognition based human detection has attracted increasing attentions since 2005. It also promotes the development of other object detection tech-niques. Researchers propose many feature extraction methods and classifier learning methods during this period. However, nowadays it is still an active research area and a challenging problem due to object’s changes on appearance and occlusion. Besides, in terms of surveillance environment, various camera settings, cluttered background and illumination changes also affect the performance of object detection systems. After reviewing various object detection methods, two new feature extraction methods are proposed in this dissertation.1) a new local binary pattern based feature Modified Symmetric Local Binary Pattern (MS-LBP). Local binary pattern based features are fast and accurate especially for human detection task. The proposed feature is a kind of modification of local binary pattern, but it focuses on horizontal property of the image and captures the characteristic of gradient based features and local binary pat-tern based features simultaneously. Thus it is more suitable to describe the symmetric property of human image. MS-LBP is embedded into the framework of Boosting and cascade to obtain high performance on detection accuracy and speed. On top of this, two feature combination methods are presented to further improve the performance of this detector. Detailed experimental results also show the importance and effec-tiveness of feature combination and dimensional reduction methods. The final trained classifier outperforms benchmark method in terms of accuracy and speed.2) a new co-occurrence matrix based feature Co-occurrence Orientation of Gradient Magnitude uniform Local Binary Patterns (CoGMuLBP). Recently, gradient orientation based co-occurrence matrix has been adopted to describe human object which achieved better performance than benchmark method. However, the gradient orientation is sensitive to noise. In this dissertation, a new feature is proposed to further improve the per-formance of object detection. It adopts local binary pattern to replace the gradient orientation and builds feature based on this orientation co-occurrence matrix. We also proposes two extensions which apply gradient magnitude and variance to weight the co-occurrence matrix. This feature achieves high performances on human detection and car detection.2) In the field of object classificationBoosting is a widely used ensemble based learning algorithm. The strategy of Boosting is to progressively combine weak classifiers to form a strong classifier with better performance. As a general and effective learning scheme, Boosting has been deeply studied in theory and has been adopted to deal with a wide variety of problems in pattern recognition. However, in order to obtain a good classifier with high perfor-mance, Boosting needs a large amount of training samples. Researches have proposed many approaches to improve the performance of Boosting. For example, different loss functions are applied to train new Boosting algorithms and different kinds of data (e.g. unlabeled samples) are taken into account to improve Boosting. In the early version of Boosting algorithms, weak classifier selection and strong classifier learning are linked together. It has been demonstrated that decoupling these two processes can provide more flexibility for training a better classifier. It is motivated by this valuable outcome, a general training Boosting-like algorithm framework is proposed. This framework s- elects weak classifiers based on class separability. It is different from conventional Boosting algorithms which select weak classifiers giving the smallest training error. Lots of subspace learning methods can be embedded into this framework to generate different algorithms. Localized Linear Discriminant Analysis and Semi-supervised Lin-ear Discriminant Analysis are adopted as examples of this framework to deal with small sample size problem. As shown by our experimental results, the proposed methods can obtain superior performances over their supervised counterparts and conventional Boosting algorithms.