| Simultaneous Localization and Mapping(SLAM)breaks through the problem that the positioning accuracy of generalized signal sources is not high in complex and changeable outdoor environments,and has become a key technology in the field of outdoor positioning and navigation of mobile robots.At present,Visual SLAM has the advantages of low cost and rich data information compared to Laser SLAM,but the traditional pure vision SLAM has low positioning accuracy in fast motion and low texture scenes,with the blowout development of deep learning technology at home and abroad,applied to a variety of hot research fields,through this technology,SLAM researchers began to research and innovate in visual SLAM.Therefore,based on the deep learning algorithm,this paper proposes a visual SLAM for constructing a semantic map of low-speed unmanned vehicles in closed parks to locate and navigate semantic maps for low-speed unmanned vehicles in closed parks.This article mainly has the following innovations:Firstly,in view of the problems of incomplete viewing angle,incomplete timing information,and poor camera field angle,which leads to only local space and semantic information between multiple cameras,this paper sets up 6 wide-angle cameras for lowspeed unmanned driving,6 cameras internal and external parameters and field of view are integrated,the field of view of 6 cameras is integrated into 360° surround view,driving around low-speed unmanned vehicles,and external parameter calibration is carried out with the vehicle IMU sensor.Secondly,aiming at the problem that the map feature points constructed by pure visual SLAM are sparse and cannot be well clustered to obtain semantic information,this paper proposes a semantic segmentation method based on deeplabv3+ neural network model optimization,and makes a proprietary thematic dataset(a total of 5600sheets)of closed park surface information,and compares it with Baidu Apollo’s public lane line dataset for training and learning.In the test and verification process,according to the special situation of the dataset,the size of the hole convolution of the network model is optimized,so that the learning effect is significantly improved,and the average intersection and union ratio is the highest 78.2%,and the experimental results show that the improved neural network model and self-made dataset are more suitable for complex surface information feature extraction.Finally,aiming at the problem of positioning and navigation of unmanned vehicles in the overall closed park,a semantic vision SLAM scheme based on deep learning is proposed,combined with the IMU sensor data equipped by the unmanned vehicle,and integrated into the ROS(Robot Operation System)platform in the Gazebo simulation software for simulation experiments,the test results are good,and the overall software system is deployed to the low-speed unmanned vehicle experimental platform to test the feasibility of the overall system in the closed park.The experimental results show that in the area with complex real environment and dense feature points,the positioning and navigation state is good,and in the case of sparse feature points,the unmanned vehicle has a small probability of localization drift.After sorting out the experimental results,the visual scheme,lidar and GNSS are fused together,the state threshold is set,GNSS is given the initial pose,and the visual system assists the positioning when the lidar positioning and GNSS signal threshold is exceeded and navigation.The test results show that when passing through the building density and turning section,the effect is poor when only lidar and GNSS are used,but the effect is significantly improved after the visual system assists positioning.In summary,this paper proposes a neural network-based semantic vision SLAM,which can run at low cost and have strong robustness under good environmental conditions.In the case of a bad real environment,it can play a strong auxiliary role. |
