Dense Elevation Mapping And Place Recognition

Dense elevation mapping is important to mobile robots’ applications such as alien exploration and jungle trap.However,dense mapping suffers from large data and high computation,which gives a huge challenge for large-scale map storage and real-time map update.In this thesis,we propose a method of local dense mapping with traversability for real-time obstacle avoidance of mobile robots,a method of scalable and globally consistent dense mapping to meet the return navigation requirements,as well as a place recognition fusion network based on a 2.5D dense elevation map representation.The main contributions are as follows:1.A real-time mapping system for constructing local dense maps with traversability is proposed.The mapping system uses the 2.5D elevation map as the map representation,which is updated by the sensor data and the robot’s poses.The system introduces the evaluation of the traversable domain to meet the requirements of obstacle avoidance and focuses on improving the algorithm structure to achieve efficiency.Compared with the original method,the local dense mapping using the acceleration strategy is faster than 10 times and runs at 7HZ on Tx2.2.An online mapping system for constructing a large-scale and globally consistent map is proposed.By decoupling the local and global mapping threads,the mapping system can output both local dense elevation map and global dense elevation map.The global dense elevation map is constructed as a series of submaps and is deformed to achieve global consistency according to the optimized poses.In the meanwhile,the map storage will not increase indefinitely in the bounded environment through the map maintenance mechanism.Experimental results on both simulated and public datasets validate the high accuracy,efficiency,global consistency and scalability of the proposed global mapping system.3.A place recognition network using dense elevation maps with hierarchical fusion of texture features and structural features is proposed.This paper presents a deep neural network for place recognition fusing visual images and elevation map information,which inherently aligns the camera and LiDAR according to the internal and external parameters of sensors,yielding a semantic feature representation with sensible geometry.On the Oxford dataset,our network achieves the best Recall@1—88.93%,which outperforms the other state-of-the-art methods.In the meanwhile,the network also obtained the best results in generalization evaluation on the cross-city scenario dataset.