Laser And Vision Data Fusion For Large-scale 3D Scene Reconstruction

The simultaneous localization and 3D scene reconstruction is the hotspot and focus in the field of intelligent robot.The vision sensor has the characteristics of high sampling frequency and can collect the color and texture information of the scene,but it is easily affected by environmental factors,such as light.The 3D laser can obtain the scale information of the scene and is not affected by environmental factors such as light,but its data acquisition frequency is slower.So this paper utilized monocular camera and three-dimensional laser to simultaneous localization and 3-D scene reconstruction in unknown environment.In order to find the joint calibration parameters,this paper presents a joint calibration method which based on image corners and point cloud corners pairing.The method uses a scattered point cloud to obtain a non-distorted point cloud reflection image through perspective projection,and then establish a matching pair between the reflection image corners and the visual image corners.Finally,this paper use the Gaussian Newton iteration method to solve the optimal solution to get the joint Calibration parameters.In order to realize the simultaneous localization and 3D scene reconstruction,this paper use the IMU to compensate the 3D point cloud data,then do the registration between the point clouds.In order to improve the efficiency of 3D reconstruction,this paper solves the real-time problem of mass cloud registration based on ICP algorithm: this paper first use feature points to do the registration between the point cloud after extract the corner points and plane points;then we will register the already calibrated laser point cloud data with historical cumulative point cloud,and we will get more accurate odometry data and 3D reconstruction data.In this paper,to verify the validity and accuracy of the algorithm,the algorithm is tested by using the DUT datasets,the EQRC datasets and the KITTI datasets.In order to solve the problem of registration failure when the 3D laser turns quickly,we propose that utilize characteristics of visual sensor high sampling frequency to calibrate odometry.The method extract the Harris corners of the image,then use the KLT algorithm tracking corner points.We provide initial calibration for laser point cloud data through the fusion of laser scale information.Experiments show that this method can effectively solve the problem of registration failure when the 3D laser turns quickly.Finally,this paper uses the open source tool g2 o to make the back-end optimization of the closed-loop experiment.