Research On Autonomous Following Technology Of Unmanned Platform In Complex Environment

As an application of machine intelligence,the autonomous target follow behavior of unmanned vehicles can be applied to various task scenearios.This paper studies the key technologies of autonomous target follow,including target recgonition and localization,and motion control with obstacle aviodance.The main research contents and innovations of this article are as follows:To address the consistency problem of target recognition in autonomous following tasks in complex environments,a feature representation based on sparse point cloud appearance model is proposed,which converts the problem of target recognition in autonomous following tasks into a target-non-target binary classification problem.A binary-classifier is trained based on the integrated learning method to be applied to the tracking process to identify the target in real time and the position in real 3-D world is tracked.This method can be applied to the target recognition and positioning module of autonomously following unmanned platforms with time efficiency and robustness.For the reason that lidar point cloud data is easy to be occluded and difficult to recover due to the lack of a unique description of the target after occlusion,a method for identifying,tracking and positioning target personnel based on visual features and binocular depth images is proposed.Based on the deep learning method,the target person detection and recognition network framework is designed to solve the problem.By training the backbone network extracted from the visual feature of the target person,the depth feature of the candidate target person is extracted online during the tracking process,and the distance measurement with the template is performed to select the final confidence target.The attention mechanism is used to filter out unlikely targets,improve its calculation efficiency.The template is updated at a certain rate during the tracking process.In the final recognition result,by combining the segmentation mask and the binocular depth image to calculate the center of gravity of the target for positioning and tracking.Experiments show that,in the tracking process,this method can solve the common target recovery after occlusion in the autonomous follow task,and obtain stable target positioning results,which has good reference value.To address the problem of pathfinding in the process of autonomous following,a motion control and obstacle avoidance method is proposed.Based on the vector field histogram method,real-time local path planning in the process of following is carried out,and a control method for autonomous following is proposed.The experiments of different following behavior under different following distances were carried out both in the simulation environment of the robot operating system and in field,which proved the effectiveness of the method.Based on the research of single target following technology,the multi-vehicle formation control is studied.To address the problem of false targets in complex environments,based on the simplified method of point cloud positioning,combined with the ultra-wideband distributed positioning system and millimeter wave radar detection results,the target-level fusion of the autonomous recognition results of each sensor is performed.The system's overall design and architecture is given.Experiments results show that by autonomously following the pilot vehicle,vehicle formation control can be achieved,which has good reference value.