Research On Target Following Method Based On Lightweight Target Detection Algorithm And Learning Human Controller

PTZ target following technology,that is,the PTZ is equipped with imaging equipment,and adjusts the orientation of the PTZ according to the movement trajectory of the specified target in the field of view,keeping the target within the center of the field of view.The main contents of research are as follows: target detection and PTZ motion control algorithm.With the development of artificial intelligence,the research results of target detection and identification through visual sensing methods emerge in an endless stream,but usually the neural network structure is complex and the calculation amount is large,which does not meet the needs of practical applications.For PTZ motion control algorithms,most of the research is based on classical control theory.In the face of high-complexity motion trajectories,the control algorithm with the same parameters cannot effectively keep track of any complex motion trajectories,that is,poor generalization.Aiming at the above problems,a target following method basing on a lightweight target detection algorithm and a learned human controller is proposed.Basing on the YOLOv5 s target detection algorithm,it is designed in three aspects:reducing algorithm depth,reconstructing algorithm structure,and pruning algorithm redundancy,and a lightweight target of detection algorithm are finally realized.The algorithm depth coefficient is designed from the original 0.5 to 0.4.The feature extraction part and inference part of YOLOv5 s are then reconstructed by using depthwise separable convolution and Ghost module.Then,sparse training and model pruning are performed on the improved detection algorithm in the first two steps.After experimental analysis,the improved lightweight detection algorithm is better than YOLOv5 s.The processing speed on the same CPU is doubled,the model size is compressed to one-tenth of the original size,and the amount of floating-point calculations is compressed from the original 16.5 billion times per second to 1.2 billion times.The detection accuracy evaluation index of the proposed algorithm,Mean Average Precision(m AP),is not significantly different from YOLOv5s(only decreasing by 1.82%).By applying the learning human controller to the motion control of the PTZ,the control algorithm with the same parameters can keep a good follow to any complex motion trajectory,and the structure is simple and the calculation efficiency is high.After analyzing the position error input of the target following problem,the control algorithm is designed by using the learning human controller kernel-Extreme Learning Machine(ELM).Using Lyapunov’s stability theorem,the constraints to ensure the stable operation of the control algorithm are analyzed and applied to the training process of ELM.Using MATLAB simulation,the training data for ELM will be obtained and trained.Self-designed two-axis gimbal,combines with the above lightweight detection algorithm to collect the actual motion trajectory.The proposed control algorithm is experimented with the simulated motion trajectory and compared with the sliding mode controller and the PID controller.After experimental analysis,the proposed learning human controller has a unique advantage in generalization,and the following performance is better than the other two controllers in the face of different motion trajectories under the condition of keeping the parameters unchanged.