Research On The Moving Obstacle Vision Avoidance Based On Two Demensional Local Map

In the field of mobile robot,study on static obstacle avoidance under known environment is sufficient,but practically mobile robot usually work in unknown environment with moving obstacles,so mobile robot takes advantage of the sensor to obtain information about the distance between obstacle and itself,as well as the size and speed of the obstacle,and then adjust its linear and angular speed to avoid the obstacle.This paper uses RGBD camera to capture point cloud on which a cost map is built,then uses dynamic window approach to compute safe route of mobile robot to avoid obstacles.First,the data from RGBD camera is preprocessed.The relation among camera coordinate,camera image coordinate,and camera pixel coordinate is derived,and then both rgb and depth cameras are calibrated.Depth image captured by depth camera is used to get point cloud,and then the reasons why point cloud become noisy are explained,after that methods of filter and clustering are used to decrease point cloud noise.Floor plane in point cloud is detected using Ransac algorithm,and then small and suspended obstacles are detected by point cloud clustering.Second a method to detect moving obstacle is proposed.First of all,optical flows are calculated by Lucas-Kanade method,then optical flows that have high LOF value are removed.For optical flows that located on a plane,they can be rectified by plane homography matrix.For a point in the scene,its optical flow orientation is independent of its depth to camera,so a optical flow orientation model that describes all optical flow orientations belong to static objects can be built.Part of static background can be identified by point cloud process result,and then parameters of optical orientation model can be computed using optical flow belongs to the detected part of static background.If the orientation of an optical flow computed by optical flow orientation model is different from what computed by Lucas-Kanade method,this optical flow belongs to moving object,so that point cloud of moving object can be segmented from all point clouds.Finally the position and speed of moving object relative to local avoidance coordinate are computed using transform between robot coordinate and local avoidance coordinate.Third the dynamic window approach is used to avoid obstacle based on the local cost map.First of all the data from IMU and robot odemetry is fused by an extended kalman filter.Then processed point cloud is projected to the cost map forming obstacle layer,the position and speed of moving object is used to form moving obstacle layer,addig cost to the surroundings of obstacle forms inflation layer,so that an ultimate local cost map can be built using obstacle layer,moving obstacle layer and inflation layer.Finally dynamic window approach is used to avoid both static and moving obstacles.