Moving Object Detection And Tracking Algorithm Research Based On Complicated Background On Earth

Moving object detection and tracking is a key technology in computer vision and a hot research topic in optical imaging system. Novel ideas and theories are continuously introduced to this topic in recent years, which forcefully promote the technology progress, bringing in a lot of new image processing methods for moving object detection. The paper focuses on the moving object detection and tracking in images with complex backgrounds on earth. The main work and achievements of the paper are as follows:1?According to the inaccurate segmentation problem of the optical flow field, a new optical flow field segmentation method based on the speed accumulation of the same direction is proposed, which provides foreground knowledge to the next segmentation. The basic optical flow equation based on the fractal Brownian motion model is deduced to determine the flow of natural images. Considering the two-dimensional vector of the optical flow, we calculate the sum of the speed of same direction to form a histogram of weighted speed direction. The direction of the histogram peek is the direction of the moving target. Then, we can segment the moving target from the complex backgrounds, which provide positive samples to the next work.2?According to the shortcomings of the traditional optical flow, which leads to the fail of detection in complex backgrounds on earth, a new detection method based on foreground and background knowledge two-stages segmentation on visual saliency map is proposed. The SLIC algorithm is used to divide the image into several super-pixels, and the ranking manifolds algorithm calculates the saliency price of every super-pixel, which makes of the saliency map. Also, the four sides of the image are regarded as the background, which are negative and the result of optical flow method are positive. At last, we use all the labeled super-pixels and ranking manifolds to give the new saliency price to all the super-pixels. Then,we can get the moving target using the easy threshold segmentation algorithm.3 ? According to the shortcomings of TLD(Tracking-Learning-Detection) algorithm in occlusion and influence of the similar target. the CTLD(Context Tracking-Learning-Detection)algorithm is proposed. we extract the Haar-like feature of the target and use Random Measurement Matrix to compress the high-dimensional Haar-like feature space to low-dimensional. The SVM based on structural risk minimization is used to learn and train the low-dimensional feature and find the most same sample as the result of SVM, which is the maximum distance from the hyperplane. Context learning algorithm acts as the tracking machine to track the target, which formulates the sptio-temporal relationships between the object of interest and its local context. We deduce the scale factor in detail and rotation factor with Mean-shift algorithm. At last, the result of SVM and the result of tracking machine is used to calculate the NCC with the history target model collection, respectively. The final result of the method is the best of the two results.4?Apply our detection and tracking method into several image sequences and compare to the classical methods, the result show that our method has a better performance in target detection and tracking in complex backgrounds on earth.