Research On Key Issues In Quickly Image Stitching Based On Low Altitude And Light UAV Image

Due to UAVs(Unmanned Aerial Vehicle) image acquisition are typically easy, fast and feasibility, which are well suited for another fields. Now, UAVs are commonly equipped with different sensors such as compact digital cameras, SLR(Single Lens Reflex) cameras, metric cameras, light weight LIDAR sensors and line scan hyperspectral sensors. Comparison to the rapid development and growth of UAVs hardware and platform, the processing capacity of post processing software for UAV imagery is clearly insufficient efficiency, especially its real-time performance on the spot. Meanwhile, for low altitude and light weight UAV image, the poor of UAV platform posture parameters are acquired when the images are captured with adverse weather conditions, like as windy and rain. And then, with the influence of the poor stability of the platform, the image geometric distortion is sensitive to terrain elevation. Just considering the image without terrain elevation, the linear polynomial model is hard to remove the effects of image geometric distortion. When many images are acquired in the survey area, the problem of geometric accuracy is more prominent. Now, it is difficult to acquire the precise of the posture angles(pitch, yaw and roll) in the UAVs platform. So when considering to use the collinear equation to create a planimetrically correct image, we should find the methods how to calculate the external parameters.As mentioned above, for the UAV image with low altitude and light weight, in this dissertation, GPU as hardware factor is considered to enhance the processing performance, and meanwhile the algorithms are improved to reduce the consumption time. In the last, we developed the auto stitching software for UAV imagery on the spot. The main research works as followed:1. GPU-Sift feature descriptor extraction; Based on the performance of each step in Sift feature extraction and the characteristics of high-resolution images of UAV, we discussed the configuration of some parameters in Sift algorithm. Under the premise of ensuring the accuracy of feature matching, we limited the number of octave of scale space. We got the improvement: limiting the number of the feature point and reducing the consumption time. Meanwhile, we extract the Sift feature descriptor with GPU parallel computing, and evaluate the extracting time.2. Constructing potential matching image pairs; For existing the large number of invalid matching relationship in the large scale image sets, we construct quickly the potential matching image pairs with the distance relationship of using image GPS information. Meanwhile, we discussed that the image retrieval method with vocabulary tree based on the Sift feature descriptor. This method also can construct quickly the potential image pairs without any other geometric information.3. GPU-Sift feature matching; In this stage, searching the nearest distance based on the constructed KD tree are performed on the GPU and constructing the KD tree do it on the CPU.4. Ransac algorithm; Based on the 5 point algorithm for solving essential matrix, we discussed the three parameters in ransac algorithm. Though experiment on the high resolution UAV image, we got the parameters suited for UAV images.5. The relative orientation of multi-image; In this dissertation, the camera calibration parameters(include the internal calibration parameters and lens distortion parameters) are assumed known for each camera. Based on essential matrix in this case, we can recovery the relative orientation of each image. Here, we mainly discussed global bundle adjustment in which we calculated the relative orientation relationship and object coordinates in the spot.