Research On Depth Estimation And 3D Object Detection In Indoor Monocular Navigation

In indoor navigation,depth estimation and 3D object detection are two important tasks.In recent years,with the rapid development of artificial intelligence technology,low-cost monocular navigation has gradually become a research hotspot.However,monocular vision lacks spatial scale information,resulting in low accuracy of depth estimation and 3D object detection.To solve this problem,this dissertation combines the low-cost 2D LiDAR to improve the accuracy of depth estimation and 3D object detection in indoor monocular navigation.The main work and contributions are as follows:1.An indoor monocular depth estimation algorithm fused with 2D LiDAR data is proposed.Two feature extraction networks are used to extract RGB image features and 2D LiDAR data features respectively.Channel attention mechanism is used to fuse the two features.Use skip connections to enrich the details of the results.In order to verify the effectiveness of the algorithm in indoor navigation scene,a depth estimation dataset with 2D LiDAR data is made.Experimental results show that the algorithm proposed in this dissertation performs better than the existing indoor monocular depth estimation algorithms on both NYUDv2 dataset and self-made dataset.2.Based on the above monocular depth estimation algorithm,an indoor 3D object detection algorithm based on monocular depth estimation is proposed.The algorithm converts the depth predicted by monocular depth estimation into point clouds and voxels through the data form conversion module,and then uses the existing point cloud 3D object detection algorithm to detect the 3D bounding boxes.During the algorithm training,the monocular depth estimation algorithm and the point cloud 3D object detection algorithm are trained independently,the transfer learning is used to improve the accuracy of the monocular depth estimation algorithm,and then the two algorithms are connected in series through the data form conversion module for joint training.In order to verify the effectiveness of the algorithm in indoor navigation scene,a 3D object detection dataset with 2D LiDAR data is made.Experimental results show that the algorithm proposed in this dissertation performs better than the existing indoor monocular 3D object detection algorithms on SUNRGBD dataset,and can still maintain high accuracy on self-made dataset.3.In order to further verify the effectiveness of the above algorithms in indoor navigation,combined with ORB-SLAM2 SLAM system which is commonly used in indoor navigation,an experimental verification system based on monocular depth estimation and 3D object detection is constructed.The monocular depth estimation algorithm is connected with RGBD mode ORB-SLAM2.A 3D object filtering algorithm and a 3D object location optimization algorithm are proposed to add the 3D object detection results to the map.The test results show that the verification system constructed in this dissertation can improve the stability and accuracy compared with the monocular ORBSLAM2,and can also provide 3D bounding boxes semantic information.