Research On Autonomous Person Following Method Of Mobile Robot Based On Monocular Camera

With the rapid development of artificial intelligence,mobile robots have gradually developed from artificially controlled movement in the early years to autonomous movement.In the human-machine cooperation scenario,the ability of mobile robots to follow the target person autonomously is becoming an incremental demand,which not only liberates the hands of the navigator,but also greatly reduces the cost of the robot’s real-time autonomous path planning.Monocular camera has attracted more and more attention in the field of human following due to its simple structure,easy calibration and identification.The key to building a complete and stable person following system is the detection,classification and reidentification of the target person.Therefore,facing the unstructured real environment,it is of great theoretical and practical significance to study the autonomous following method of mobile robot and realize the recognition and following of target person based on monocular camera.This paper focuses on three key issues:person detection,target classification and target reidentification after loss,and carries out the following research contents:(1)According to the needs of person detection based on monocular camera,the imaging principle and calibration principle of monocular camera were analyzed,and the camera calibration was carried out.Based on the processing of the original camera data,a human detection algorithm is constructed.Aiming at the problem of camera vibration in the process of mobile robot motion,an image smoothing mechanism based on mean filtering was proposed,which effectively overcame the impact of vibration on camera receiving and processing data,and improved the stability of person detection algorithm.(2)In order to solve the problem of person classification in the following process,a person classification network based on CCF and Online boosting algorithm was proposed,which could achieve real-time classification of target person and other persons by extracting and learning the appearance features of the target.In addition,aiming at the problem of low re-identification accuracy after the loss of the target,a target recapture mechanism based on RNN is proposed.By combining the motion information and appearance features of the target,the reidentification accuracy after the loss of the target is improved.Experimental results show that the proposed person classification network and target recapture mechanism are effective.(3)In order to solve the problem that a single identification network cannot dynamically adapt to the following environment,a target continuous identification algorithm based on GWR network is proposed.Based on the proposed CCF features,the algorithm can continuously learn the target person and realize long-term and stable identification.In order to overcome the forgetting problem of network,this paper adjusts GWR network structure and proposes regularization mechanism based on neural network structure.Secondly,aiming at the problem of target loss,this paper proposes a double-layer memory replay mechanism to improve the accuracy of target continuous identification and solve the problem of target loss.The real-time results demonstrate the effectiveness of the proposed algorithm.(4)Aiming at the real application scenario,the mobile robot experiment platform is built,and on this basis,the proposed algorithm is compared with other traditional algorithms,which proves that the proposed algorithm has higher identification accuracy.In addition,the stability of long-term following is verified in the real campus environment.The experimental results show that the proposed algorithm can meet the needs of long-term autonomous following of robots,and has practical application value.