Research On Multi-sensor Fusion Navigation Method Of Four-wheel Mobile Robot On The Slope

When four-wheel mobile robot performs tasks on the slope,it needs to make real-time path planning and accurate positioning and navigation on the dynamic slope to avoid moving obstacles.Partial slipping of the wheels of the robot caused by the sloped road surface,which causes a large error in the actual motion trajectory and path planning of the four-wheeled mobile robot.In order to improve the robustness of the robot's navigation on dynamic slope,this article uses monocular vision-assisted laser for depth information fusion,the main contents are as follows:First of all,kinematics modeling and trajectory tracking of four-wheeled mobile robots on hard slopes: The causes of four-wheeled robot skidding on hard slope environment are analyzed.The kinematics model of the four-wheel mobile robot is established under the consideration of slope skidding,and the static and dynamic stability of the robot is summarized.An adaptive trajectory tracking control algorithm is proposed to update the slip parameters online for slope slip,thereby reducing the effect of slip on the pose error of the robot path planning.Secondly,multi-moving object detection on the dynamic slope based on Monocular vision: The problem of low real-time and low accuracy of traditional moving object detection algorithms in acquiring images on dynamic slope is analyzed and summarized.A feature tracking image stabilization algorithm based on optical flow is proposed.The decision tree is trained by the ID3 algorithm.The 16 pixels on the circumference of the feature points are input into the decision tree to screen out the optimal SURF feature points.An adaptive detection interval time difference sliding window object detection algorithm is proposed.Multiple feature points are extracted from the reference frame.The time difference sliding window with 3 frame intervals is used for background difference.In order to compensate the background motion with differentproportion,we track multiple moving objects across multiple frames with interval time.Therefore,the running speed and accuracy of the multi-moving object detection algorithm are improved.Thirdly,dynamic path planning based on deep reinforcement learning on the slope: In order to improve the time efficiency of path planning on the dynamic slope,the DDQN(Double Deep Q Network)method and the tree structure method are combined.The path information is stored in the optimization tree structure,and the incomplete path and the excessive path are discarded.Making the TDDQN(Tree-Double Deep Q Network)algorithm select the best action in the current state after performing fewer actions,so as to obtain candidate paths that meet the conditions.The non-maximum suppression method is used to select the best path from multiple eligible path candidates.The convergence speed and time efficiency of the path planning algorithm are improved.At last,research on navigation method based on multi-sensor fusion on slope: The monocular camera and lidar are jointly calibrated.Aiming at the problem that a single lidar is prone to produce depth information errors when navigating down a slope,a method of fusing depth information with a monocular camera and lidar is proposed.The fusion of depth information through the laser point cloud on the foreground target reduces the depth information error of the detection object.The necessary key frames are added when turning,which improves the robot rotation accuracy.Simulation and experiment were carried out under ROS(Robot Operating System)environment.Compared with Cartographer SLAM,the navigation translation and rotation accuracy are higher.The robustness of the proposed method is verified.