Research On Application Of NCS-based Neural Network In UAV SLAM Field

During the development of UAV,safety is an eternal topic.Obstacle avoidance technology is realized by various sensors,which is an important cornerstone to ensure the safety of UAV.However,limited by the load and volume of UAV,it is poor on onboard sensor and computing platform in performance.The sensors can only realize the basic perception and understanding of the outline of obstacles,it cannot choose a reasonable safety distance for different types of surrounding obstacles.This greatly limits the scope of UAV application in complex environments.In order to make UAV understand the surrounding environment and improve the ability of avoidance.In this paper,a semantic SLAM system for UAV is designed and implemented based on Neural Compute Stick(NCS).Compared with the traditional visual SLAM system,the semantic SLAM system proposed has been improved in the following two aspects.On one hand,the increased semantic segmentation network module is based on convolution neural network,which uses NCS to achieve the semantic segmentation network deployment on UAV platform,and improves the traditional semantic segmentation network by depth information,the network structure and data loading method according to the performance characteristics of UAV usage scenarios.On the other hand,the mapping method of the traditional visual SLAM system is improved.Firstly,semantic map is obtained by combining the point cloud data from the SLAM algorithm with the semantic information from the semantic segmentation module,which is used to set safety for surrounding objects;Secondly,the octree mapping technology is used to splice early warning map with the safety distance to avoid obstacles.In this paper,semantic segmentation module and SLAM system are verified.The experimental results show that the improved semantic segmentation module based on NCS can realize semantic segmentation of the surrounding environment on the premise of real-time.At the same time,the SLAM system can build semantic maps based on semantic information provided by the semantic segmentation module.This paper validates the effectiveness of the system with PX4 flight control obstacle avoidance by semi-physical simulation.