Design and Implementation of UAV Vision Tracking System

UAV is increasingly popular in the field of military, civil and scientific research, since its low loss, strong adaptability and high mobility. This paper introduces the design and implement of the visual tracking system based on the rotor UAV, and gives the UAV a function of continuous and stable tracking of the ground target. Based on the standard Mean Shift algorithm, we introduce the improved object tracking algorithm with the spatial histogram and scale estimation. The robustness to scale adaptive and background related interference is improved with the enhanced tracking algorithm. It is finally implemented in our image processing board with TMS320C6657 and ZYNQ 7020. The whole system is optimized and accelerated by both of the two core of DSP. We finally achieve the real-time tracking and monitoring of the ground target in the range of 100m ~ 1km. Introduction In recent years, UAV is more and more popular as for its god perspective. And it is more usually used in the field of computer vision applications. In this paper, we choose the rotor UAV as the experiment platform. And a image processing board with TMS320C6657 and ZYNQ 7020 is designed and produced too. That let us to give the UAV a function of continuous and stable tracking of the ground target. We use the standard Mean Shift target tracking algorithm as the visual processing basic algorithm in this system. The improved object tracking algorithm with the spatial histogram and scale estimation is proposed to make the algorithm more robust. Finally, the proposed tracker is implemented in the image processing board, and it completes the real-time solution of the target position in the image sequence. The UAV achieves the real-time tracking and monitoring of the ground target in the range of 100m ~ 1km with our achievement. Mean shift tracking is a popular algorithm for object tracking since it is easy to implement and it is fast and robust. It was first proposed by Fukunage and Hostetler in 1975 [1]. And it was first applied in the target tracking field by Comaniciu. The Mean Shift is a non-parametric mode-seeking method for density functions with no prior knowledge. The original Mean Shift algorithm uses a color histogram as a feature description of the tracked target. And the tracking window is fixed. Once the target is zoomed in, the rotation and the target color are similar to the background, the tracking performance is degraded or even lost. For the problems of the Mean Shift algorithm, the tracking window is scaled at a fixed scale of ±10% was proposed by Comaniciu, and that the window can be adjusted by multiple iterations [2]. However this method cannot be better adapted to the process of getting bigger. Collins uses the Lindeberg scale space theory to adapt to the target scale change [3]. Liang et al. search for the object boundary by correlating image with the four templates [4]. This kind of method often has the problem of large computation and difficult to transplant, and it is difficult to realize real-time processing in the embedded platform. Zhao et al. exploit affine structure to recover object relative scale from feature point correspondences between consecutive frames [5]. Methods depending on feature matching are able to estimate the scale quite robustly, but they cannot be 1539 5th International Conference on Frontiers of Manufacturing Science and Measuring Technology (FMSMT 2017) Advances in Engineering Research (AER), volume 130