Research On Key Techniques Of Efficient Cloud-Terminal Fusion Based SLAM Algorithm

As the fundamental method for robots to achieve autonomy,simultaneous localization and mapping(SLAM),is limited by the computational storage resources carried by local capabilities and faces the challenge of task processing performance degradation in scenarios with higher timeliness and scale requirements.At the same time,with the rise of cloud computing technology and the concept of ”cloud robot,” using the capabilities of cloud platform to improve the task efficiency of robots,is considered a promising way to solve the related problems.However,in the large-scale and complex task scenarios,to enable the cloud platform and the robot to perform SLAM task processing efficiently,the existing work still faces enormous challenges.The most outstanding one is the lack of adaptability in existing computation offloading methods when entering dynamic and scaled scenarios.The unstable of networks and deterioration in QoS in the existing cloud platform can have a significant impact on the task processing capability.Another one is that the cloud platform-based method is hard to utilize shared data efficiently.Although cloud platform gives the SLAM the ability to share map data among multiple robots,there is no practical way to integrate with these local sub-optimization SLAM process.Based on the background of cloud computing and autonomous robot SLAM,this thesis explores the above-mentioned critical challenges in the SLAM problem and focuses on the topic of “the way to make existing SLAM algorithms and applications efficiently utilize resources in cloud platform.”The main work of this thesis is detailed in follows.1.Focusing on the difficulty in existing robot applications to be deployed into cloud platforms,this thesis proposes a transparent service encapsulation mechanism.It aims to give robot applications the ability to transparently offload computing modules to cloud platforms,and keep communication with the local robot in real time without any code modification.To meet the service quality and stability required for robot applications,a coordinated service quality assurance mechanism is designed on both the cloud platform and the robot side,to cope with performance jitters such as network failure and QoS degradation in cloud.2.To tackle the communication performance bottleneck in multi-robot collaborative SLAM,this thesis formalizes the computational and communication models of peer-topeer based multi-robot architecture.With comparison and analysis on the performance of local native in computation offloading,we abstract the communication bottleneck mode in three different levels.For each different communication bottleneck mode,different optimization mechanisms are proposed in each aspect.Through these methods,the promoted framework achieves the goal of dynamically optimizing on the data transmission path and transparently reducing the redundant messages between the robot and the cloud platform.3.To solve the problem of insufficient scalability of existing multi-robot collaborative SLAM,this thesis further proposes a method for multi-host parallel processing of SLAM tasks to improve the number of supportable robots.In communication optimization,the proposed multi-tool algorithm groups robots within the cluster and highbandwidth low-latency communication modes for different message types;Regarding consistency,to meet the sharing and delay requirements of different data types,a lineagebased consistency maintenance method is also proposed.4.To efficiently utilize the shared maps among robots in the cloud platform,this thesis takes loop closure detection and relocalization in SLAM as the research object and proposes methods to effectively integrate map sharing with existing SLAM algorithms.Based on the spatial pose model in loop closure detection,an embedding detection optimization algorithm based on spatial adjoin relation is presented.This algorithm aims to improve the problem of too few matching in the process of detecting loops in SLAM method which has memory constraint.Also,a relocalization candidate expansion method based on cloud platform sharing is proposed correspondingly to improve the positioning speed of the existing relocalization method.The materialization of the above work is finally implemented as the prototype Cloudroid/-Cloudroid Swarm offloading framework and the cloud-based algorithm.The former is based on the proposed computation and collaborative offloading method,and the latter provides A specific optimization strategy in efficiently utilizing the cloud platform to assist the SLAM task.This proposed method and implemented prototype have been evaluated in a series of SLAM experiments.And the efficiency is demonstrated by the efficiency in a number of real-world scenarios and typical SLAM methods such as ORB-SLAM,RTAB-Map and RGBDSLAM.