Research Of Method In Human Detection Based On Second Generation Bandelet Transform

Human detection has attracted a lot of research interests in recent years, because it has several important applications in computer vision, e.g., video surveillance, smart vehicles, human-robot interaction and content-based image filtration. However, human has been proven to be a much more difficult object to detect, because of the wide variability in appearance, due to clothing, articulation and illumination conditions that are not common in outdoor scenes. In this paper, based on the second generation Bandelet transform, we followed the geometric flow of images and presented a new feature extraction method to improve the accuracy of human detection in images.Besides, many researches had been carried, focused on reducing time cost, improving the accuracy of region segmentation and enhancing the robust of detection. It involved the method of optical flow, image segmentation and the human detection method based on body-parts. Specifically, the paper proposed three original human detection methods based on second generation Bandelet transform.1) Based on the second generation Bandelet transform, we followed the geometric flow of images and presented a new feature extraction method to improve the accuracy of human detection in images. Here the Bandelet coefficients and their statistical values were extracted as the features of human images, combined with linear SVM classifier, then we can classify and detect human in images.2) Different from most of the previous work, we detect motion human by region segmentation and classification through machine learning. In our method, based on optical flow, region segmentation is carried firstly and then, based on geometric flow, Bandelet transform is used to do feature extraction and classification. Some treatments were carried after optical flow field computation to de-noising and some improvements in Bandelet transform were used to reduce time cost of feature extraction.3) The third method focused on enhancing the robust of human detection by using body-parts and feature of Bandelet transform. Experiments showed that this algorithm could be an effective way to detection human in static or motion images, and it is worthy that we should do more effort into this in the future research.