| With the rapid development of artificial intelligence,target tracking using camera sensors has become possible,so the situations of people following by mobile robots using visual SLAM(Simultaneous Localization and Mapping)technology based on camera sensors are gradually increasing,and the research value of related topics is also getting higher.In the process of people following,robots usually need to perform visual SLAM localization and mapping,and simultaneously complete target detection and continuous tracking of the followed people.Among the existing solutions,there are problems such as low accuracy and stability of visual SLAM algorithms;general visual SLAM maps do not contain semantic information leading to poor adaptability;and poor stability of target tracking algorithms due to target occlusion and other factors.To address the above problems,the following work is specifically carried out in this paper.(1)In order to improve the accuracy and robustness of localization and map building of mobile robots in unknown indoor and outdoor environments,this paper adds the wheel encoder to the camera sensor,designs a visual-wheel tightly coupled odometer based on MSKCF(Multi-State Constraint Kalman Filter)sliding window filter,and a plane constraint is added for the characteristics of the odometer.The simulation and experimental validation on KAIST dataset show that the visual-wheel odometer combines the advantages of rich visual camera information and high accuracy of wheel encoder measurement,which improves the accuracy and stability of visual SLAM during the people following process.(2)In order to solve the problems of sparse point clouds and lack of semantic information in traditional visual SLAM maps,this paper adds target detection to the visual SLAM system and designs a semantic map construction algorithm based on YOLOv3 deep neural network.Compared with the traditional visual SLAM,the semantic map can preserve the environment semantic information,which can make the map building more robust and create conditions for mobile robot navigation,obstacle avoidance and other advanced tasks;it also helps to improve the accuracy of people detection in the subsequent tracking process.(3)To solve the problems of poor tracking stability caused by target occlusion and target deformation in target tracking,this paper obtains the tracking target scale through the depth information of the depth camera,applies the techniques of lightweight dense sampling,ridge regression analysis,and scale pools to design a kernel-related target tracking algorithm based on depth information to adapt to scale changes.It is proved that the algorithm can significantly increase the stability of people tracking through people following experiments on a trolley. |
