| 英文摘要 | Object detection is one of central themes in computer vision, which describes an image in terms of the meaningful objects that comprise it, such as vehicle, humans etc. It is the basic premise for computers to further process the images and to interprete the real world. Meanwhile, object detection is the basis of other computer vision tasks, tracking by detection, for example. Besides, object detection also plays an important role in related disciplines, for instance content based image retrieval, video surveillance, information security, and so on. Generally speaking, research on object detection usually involves three issues, suitable feature expressions, effective object models, and efficacious algorithms. In this thesis, an intensive study of object detection in images are made. The main contributions of this work are as follows: 01 Propose a method of boosted forest with HOG features for human detection, and with LBP features for motorbike detection. The proposed detection method associates the random decision trees as weak learners within the framework of Adaboost. Accordingly, these random trees are dynamically combined into a strong classifier, i.e., a boosted forest. The boosting process avoids the blindness and casualness of the tree selection in typical random forest algorithm. Besides, potent features are estimated and chosen in the process. Also, the drawback of random forest is pointed out in this work. Specifically, training examples are sampled by different trees. However, the dimension is not reduced. Meanwhile, fewer samples with same dimension make classification more difficult. Thus, a kernel forest method is intro- duced, to solve the classification problem with fewer samples, especially in the bottom nodes of trees in forest. Experiments on PASCAL VOC 2008 dataset demonstrate the effectiveness and efficiency of the proposed method. 02 Present a cascade template matching framework for object instance detec- tion. Specially, we propose a 3-stage heterogeneous cascade template match- ing method. The first stage employs dominate orientation template (DOT) for scale and rotation invariant filtering. The second stage is based on lo- cal ternary patterns (LTP) to further filter with texture information. The third stage trained a classifier on appearance feature (PCA) to finally reduce false-alarms. The cascade template matching (CTM) can provide very low false-alarm-rate compa... |
修改评论