Research On Multi-UAV Cooperative Target Tracking Technology Based On Visual Perception

Target tracking is a crucial content of UAV application research,which has been much concerned by domestic and foreign researchers.Target tracking systems based on UAV platforms usually use onboard cameras to collect images,identify and locate the target based on images,and then guide UAVs to track the target.However,due to the complexity of the scene in the images collected by UAVs and the relative motion between UAVs and the target,the target may be lost in the collected images,causing the failure of the tracking task.To solve the above problems,we propose a multi-UAV cooperative target tracking technology based on visual perception.On the one hand,the visual target tracking algorithm Siam RPN is improved through the depth correlation metric,enhancing the success rate and accuracy of visual tracking.On the other hand,improve the target tracking performance by using UAVs collaboration.The main contents of this paper are as follows:(1)Aiming at the problem that Siam RPN cannot effectively distinguish similar semantic objects and lacks long-term tracking capabilities,a siamese network visual tracking algorithm based on deep association metric is proposed.Using a deep target verifier and a multi-track association tracking mechanism to convert the single-track tracking problem into a multi-track tracking problem of the target and its similar semantic objects,and improving the algorithm’s ability to discriminate similar semantic objects.To improve the long-term tracking ability of the algorithm,the target is repositioned by using the global target relocation strategy when the visual tracking error occurs.The AUC of the algorithm on the OTB2015 dataset and UAV20 L dataset are 0.661 and 0.639,which are 3.3% and 11.3% higher than Siam RPN,indicating that the algorithm has better short-term tracking performance,and it has better tracking performance on long-term tracking tasks.(2)Aiming at the positioning error caused by single UAV target recognition and the inconsistency of positioning caused by multi-UAV collaboration,a distributed square root Cubature information filtering algorithm based on hybrid consistency is used to estimate the target state cooperatively.The simulation results show that the algorithm can fuse the Observational data from multi-UAV to estimate the target state so that multi-UAV can obtain more accurate and consistent state information of the target.(3)A path planning algorithm for multi-UAV cooperative tracking based on decentralized quadratic model predictive control is proposed,by maximizing the Fisher information matrix and forming the optimal multi-UAV coordinated target tracking observation position to achieve higher tracking performance.The algorithm uses artificial potential fields to convert the nonconvexity(optimization goals and constraints)in cooperative trajectory planning into external control input items,which used in the kinematic equation of the decentralized quadratic model predictive control to achieve the convex optimization properties of this algorithm.The simulation experiment results show that the algorithm can solve the approximate optimal observation trajectory of each UAV,and then guiding multi-UAV to track the target cooperatively with approximate optimal observation position configuration.This paper has verified the feasibility and practicability of our systems through multi-UAV cooperative tracking experiments in real scenarios.