Research And Implementation On Pose Measurement Using Double Camera Based On Object Model

Pose parameters,such as the position and orientation, are core parameters to reflect the motion state of targets. In military, navigation, aerospace and many other fields, pose parameters measurement has shown essential importance. Due to the limitation of hardware facilities and economics, strategies that combine photogrammetry and computer vision techniques become the main trend in pose estimation. Specifically, we first capture a video sequence of target with optical devices, then figure out pose parameters based on related internal and external device parameters. In practical applications, how to precisely measure target's pose in aerial data has become an urgent problem to be solved.Existing pose estimation approaches based on photogrammetry could be divided into single-camera vision measurement and multi-camera vision measurement according to the number of optimal device(camera) used to tracking target, or correspondence-known approaches and correspondence-unknown approaches according to the awareness degree of the correspondence between controlling points and imaging points. However, there is a limitation on the precision of single-camera vision measuring results, because only one station could not provide much enough information. While the correspondence-known approaches usually become impractical since they require controlling points be specified manually beforehand. In this paper, we focus on designing a double-station target pose parameter estimation method with unknown correspondence between controlling points and imaging points. After a summary study of existing methods, we provide with a model based double-camera target pose estimation method.In our method, neither do we specify target controlling points beforehand, nor do we perform lots of machine learning to train template library. We first separate foreground targets by image segmentation for images in the two stations, respectively. Then we employ contour extraction and extract contour features that reflect the global feature of the targets. Next, we generate simulated projection for target model, and perform similarity matching between the simulated image contours and the extracted contours. Next, we construct an energy function by combining the matching results in the two stations under a weighted strategy. At last, we solve this optimization problem to figure out the target pose parameters with the help of a double-station space intersection algorithm.In this paper, we spend most time on the research and analysis of two key stages in our algorithm, image processing and the optimized pose parameters solving.In the image processing stage, we study several classical algorithms in image segmentation and contour detection. We first make a detailed analysis on the advantages and disadvantages of an image segmentation algorithm, Grab Cut. Based on this analysis, we implement a continuous image segmentation algorithm, which is suitable for video sequence data. This continuous image segmentation algorithm can enhance the automation in continuous image segmentation without sacrificing effectiveness, and does not introduce too much manual intervention. For contour detection, we mainly focus on Canny. To optimize the extracted contour, we introduce template filtering, morphological processing, Douglas-Peucker polygon fitting, linear interpolation and some other algorithms.In the optimization stage, we make full use of double-station information to improve the accuracy and convergence speed of our algorithm. We first define a method for contour similarity measurement, then we employ a weighted strategy to combine the contour matching results in the two stations to construct the energy function. Finally, a double-station space intersection algorithm is applied to determine the object space, which narrows the feasible region and thus accelerates the convergence speed of optimization.At the end of this paper, we evaluate and analysis our algorithm on simulated image sequences in Open GL and 3Dmax, Its efficiency is proved.