An Object Recognition Method And A Number Of Applications

Object recognition is currently a difficult problem in computer vision area, causing more and more interest in computer vision and attention to workers, the main process of which is detecting the interest object in the image using the relevant algorithms. Meanwhile, It's also a key challenge in many applications such as robot navigation, human-computer interaction, image understanding, image annotation, content-based image search, intelligent video surveillance and so on.. Research in this topic will promote the key issues in related applications better resolved.This paper studies on the object recognition method based on feature matching combining segmentation validation. Based on the previous work of object recognition method using voting mechanism, our approach integrates SIFT (Scale Invariant Feature Transform, the following are referred to as SIFT) feature extraction and voting process, and improves the statistical method of voting results using MeanShift algorithm. Experiment shows better results can be achieved.This paper also describes two object recognition applications in detail:content-based image search and video trackingIn the image search field, currently the mainstream technology adopted by Google and Baidu and other search engine is finding a picture through the text. Users enter text messages, the system returns the picture corresponding to the text. However, The subject of our topic is when user input an image, the system should return some similar pictures. Sivic and Zisserman refer to the tf-idf model in the text information retrieval. Experiment shows good results can be obtained. First, cluster on these features extracted from the training set using Kmeans, the resulting set for each cluster which is defined as a word is associated with an inverted file(which reflects the word's distribution among all the pictures situation in the training set). Then quantify the features extracted from the query image into words above. Finally evaluate the similarity between the query image and training images. Nister and Stewenius put forward a hierarchical clustering approach based on previous work. It significantly reduce the time required for quantifying features using the tree index. Our work inspired from previous work, exploiting fuzzy classification theory to improve the process of feature quantification, is implementing a small image search system.In the video track area, there's enormous challenge to track in practical application because of the complex dynamic background, scaling of the moving target, light intensity changes and so on. The paper expects to achieve a robust real-time tracking system, to overcome the scale changes of the moving target, partial occlusion, light intensity changes and other adverse factors. Our approach exploiting the object description method based on feature and regional statistic, is tracking moving objects under the framework of kalman filter whose state vector has been improved to increase the discrimination capability between background and moving target.