Research On Ground Recognition Of Sweeping Robot Based On Few-Shot Semantic Segmentation

Society is moving forward and technology is changing,Cleaning robots in smart home products are attracting more and more attention,which can improve people’s quality of life.There are many well-known brand robot manufacturers in the market,in the development of sweeping robot mostly pay attention to the path planning,road obstacle avoidance,object recognition and other technical aspects of the research.However,in the cleaning process of the sweeping robot,the cleaning strategy remains the same,which is an important challenge for the changing ground environment cleaning.Therefore,how to improve the cleaning degree of the sweeping robot on different pavement has become a problem that needs further study.The objective of this paper is to investigate the use of semantic segmentation technology for ground recognition in sweeping robot tasks.The subsequent section provides an overview of the research conducted in this paper.Above all,this paper aims to explore how semantic segmentation technology affects the ability of sweeping robots to recognize the ground.To this end,this paper firstly collected more than 2000 pictures containing various types of ground through various ways,and used Labelme and EISeg tools to annotate and preprocess the pictures,and constructed a data set containing various types of ground(such as wood floor,carpet,marble,etc.),which was repeatedly used in the subsequent network training and testing.Secondly,The ground recognition method of sweeping robot based on multilevel Transformer and lightweight multilevel sensor is studied.In addition,many segmentation methods are compared from the perspective of segmentation accuracy and segmentation rate.Multi-level Transformer can realize multi-scale and multi-direction feature extraction,preserve and utilize spatial relationships between different areas,resist and adapt to complex factors such as occlusion,shadow and lighting changes.Lightweight multi-layer perceptron can reduce model parameters and computation,and improve the running speed and efficiency.Under the same data set and hardware conditions,Seg Former and SETR methods with Transformer structure are superior to other segmentation methods.Meanwhile,Seg Former has higher FPS(39% increase)than SETR,but slightly lower m Io U(0.52% decrease).Considering speed and accuracy,Seg Former is more suitable for ground type recognition of sweeping robot,and it is improved and optimized in this paper.Thirdly,to address the issues of limited data and inadequate generalization ability in ground recognition tasks,a Seg Former-BAM network structure with few samples is proposed based on the combination of meta-learner and base learner.Among them,Seg Former,a lightweight and efficient Transformer,is used as a base learning device to enhance the recognition ability of non-ground categories(such as flower pots,trash cans,etc.).A meta-learner is used to optimize the segmentation of ground class by using a small number of samples,and then an integrator is used to suppress the interference caused by non-ground class to the ground class,so as to realize the interaction of two branches and generate smooth and tight segmentation results with segmentation edges.Experimental results show that Seg Former-BAM has higher m Io U(4.2% improvement)but lower FPS by 32.3% compared to Seg Former.