Research On Multiple Similar Ground Target Tracking And Path Prediction Of UAV

The technology of real-time target tracking and path prediction has a broad application prospect in security monitoring, intelligent transportation systems, military systems and so on. To confront with this challenge, based on the 7th IARC(international aerial robot competition) and quadrotor platform, a long-term target tracking algorithm using tracking by detection method is proposed, and on this basis realized targets path prediction.Firstly, a multi-target detector based on HOG&SVM algorithm is designed for the sake of getting the target tracking box automatically on the case of firstly tracking or tracking failure. Then a bootstrap method is used to improve detection accuracy, up to 98%. And result shows that it performance well on rotation, scale and illumination invariant.Secondly, an improved tracker based on Mean Shift algorithm is designed. As the original algorithm needs to manually select the target model, target templates are introduced to the tracker to achieve the target model automatically. Then, to solve tracking fail problems, a multiple judgment of tracking fail conditions is designed.Thirdly, a long-term multi-similar target tracking algorithm is designed. At the beginning, the detector runs every frame searching for potential targets and identification of the targets by color segmentation. After import the target template, enabling the tracker, returning the tracking status, reducing the detector’s searching frequency, and keep searching for the underlying target. For this, get the best balance between stability and real-time.Finally, a target path prediction method based on Kalman and least squares is proposed. After analyzing the ground moving target motion model, camera calibration and inverse perspective mapping, the target’s global coordinate is obtained, when target disappeared, least squares method is used to estimate the path of target. And prepare for the target searching.