| Moving object detection in image sequence is a key technology in computer vision, especially the moving object detection with background of infrared and background of visible light have been a hot topic in the field. With the gradual deepening of theoretical research, the technology has been widely used in intelligent transportation, intelligent video surveillance, robot navigation, precision-guided, medical testing and test measurements, and so on. The complexity of their backgrounds, the types of noise and interference, and the signal to noise ratio (SNR) vary with different images. Even in the same sequence, the SNR of each image is not the same. In addition, high accuracy of object detection is required in practice. But for different objects, the size and shape are fluctuated. All these features make the study of the algorithms for moving target detection in image sequence difficult. Therefore, the study of the algorithm for moving target detection is not only a great significant in practice, but also has a big challenge in computer vision and other related areas.The study moving object detection with background of infrared and background of visible light was carried out in this paper. Based on the analysis of the detection algorithm, several effective algorithms have been proposed. The main research works are as follows:(1) An idea is proposed to divide the existing algorithms of DBT into direct and indirect detection methods. The paper describes the basic principles and concepts of these two methods in detail, and then gives brief introduction of their typical algorithms.(2) The paper presents a detection method of infrared small object based on frequency-domain. First the background and some interference are removed by a high-pass filter, and then the edge was weakened by sharpen filter, next the small object was highlighted with the noise and interference filtered by using the image enhancement technology. Compared with the inter-frame difference method and background subtraction, it not only reserves more pixels for small object. but also enhances gray values greatly, has no false alarm points and overcomes the clouds interference in background of sky.(3) A method based on wavelet transformation for single frame infrared image detection is proposed. Firstly, an image was decomposed by using a wavelet transformation. Then with the abandon of the background component located the low-frequency of the image, the noise in high frequency by the wavelet analysis was removed to reconstructs the new image by three denoised high-frequency components. Finally, the image enhancement technology is used to improve small object gray scale for the further suppressing the interference. With this method, the small object can be detected precisely and efficiently, along with overcoming some interference problems that come from clouds and buildings.(4) This paper proposes two modified optical flow detection algorithms, which are fusion of other techniques. In the first one, a self-adaptive optical flow method is presented, which combined HS algorithm with blocking and smoothness. And it is proposed to overcome the shortcoming of poor anti-noise performance of the classic optical flow method. It realizes de-noising and segmentation in the original optical flow field. The detection result is almost the same as the result of manual segmentation. Furthermore, it has robustness when it comes to the complexity and status of the background. In the second one, to overcome the shortage of large amount of computation, the bilateral filtering technology is introduced into the pyramid flow method. First, a low-resolution image was obtained by the Gaussian down-sampling technology. and then the optical flow of the low-resolution image was calculated, and at last the original optical flow was obtained with the bilateral filtering upsampling. The detection results are better in this method, and Because of the using of bilateral filtering, the edge information of object is better retained.(5) An improved LK optical flow detection algorithm is proposed, which solves the problem that the singular matrix cannot be calculated. The method combines the neighborhood thought of LK algorithm with PSO algorithm, and estimates the optical flow of the object by the convergence principle of PSO. The method can better detect the object area, and does not affect the subsequent identification of a larger object. |
