Research On 3D Dense Map Construction Based On Unmanned Platform

Autonomous Navigation of unmanned platform is a very complex problem,which needs a variety of different functional modules to complete together.The three-dimensional dense map building module is a key part.The construction of 3D dense map is a hot research direction in computer vision and robotics.In order to ensure that the unmanned equipment can navigate safely or perform more advanced tasks,our mapping system must intensively cover all the environmental information in the real scene,and have enough efficiency to save memory resources while ensuring the accuracy.In addition,in order to adapt to different environments,the mapping system should be able to operate stably in indoor and outdoor multi-scale environment with excellent robustness.In order to overcome the existing problems and the above requirements,this paper proposes a method of Slam(simultaneous localization and localization)based on vision sensor The depth information obtained by vision sensor and the pose information calculated by slam are combined to get 3D dense map.The factors that affect the accuracy and efficiency of 3D dense map construction system based on vision sensor are analyzed and studied.This paper mainly completes the following three aspects of work:1.Aiming at the problem that the quality of indoor and outdoor 3D dense map construction based on vision sensor strongly depends on the quality of depth map provided,and the traditional vision ranging method is difficult to obtain high precision depth information,this paper proposes a binocular disparity estimation method based on depth neural network.On the basis of NAS(neural architecture search)for automatic search network structure,DCN(discrete convolutional networks)is integrated.Then,in order to improve the accuracy of binocular disparity estimation network,-a new adaptive weight loss function is designed.In order to verify the proposed algorithm,it is tested on Kitti data set.The experimental results show that the EPE(end point error)error is reduced by 37%and the parallax estimation accuracy is improved by 2.9 percentage points compared with the model obtained by lestereo algorithm.2.In view of the problem that the time error items between different hardware sensors cannot be dynamically updated by the visual inertia state estimator system,a time synchronization optimization algorithm based on binocular inertial navigation is proposed in this paper.The time error between camera sensor and IMU is added to the state vector to optimize the re projection error.In order to verify the performance of the algorithm,the comparison experiments are carried out on euroc data set with Venus and orbslam2.The experimental results show that the RMSE(root mean square error)error is reduced by 21 and 15 percentage points respectively.3.Aiming at the problem that traditional slam system can not only reconstruct the sparse point cloud map,but also can not directly construct the high-precision 3D dense map,this paper uses the depth map of each frame estimated by the neural network after the improved lestereo and the pose calculated by the slam system after adding the binocular vision inertial navigation time step optimization algorithm as the input to construct the hash based data Tsdf(truncated signed distance function)three-dimensional dense map of structure.In order to verify the performance of the reconstruction algorithm proposed in this paper,it is compared with the lestereo algorithm on the data collected from the public data set and the actual scene.In the measured data,the length of the chair is selected for measurement,and the accuracy of our method is improved by 67%compared with the previous method.