Research On Obstacle Avoidance Detection For Industrial AGV

As an intelligent device,Automatic Guided Vehicle(AGV)has been widely used in industrial manufacturing environments.It can handle goods independently,thereby liberating workers’ productivity and improving industrial production efficiency.At the same time,it can replace workers to handle dangerous goods,ensuring workers’ personal safety.When there are obstacles on the AGV’s driving route,if there is no efficient obstacle avoidance method,it will interfere with its normal transportation,and even cause safety accidents.However,the factory environment is variable.Some areas have simple scenes,with regular shapes and relatively uniform sizes of obstacles,while others have complex scenes,with no fixed shapes and inconsistent sizes of obstacles.Therefore,facing obstacles in different scenes,how to accurately detect their specific positions and detailed shapes is an urgent problem to be solved.This thesis designs two schemes using computer vision technology: two point cloud scales single stage 3D object detection network(TPC-SSD)for detecting obstacles in simple scenes,and a point cloud panoptic segmentation network via voxel transformer and feature completion(VTFC-Net)for segmentation of obstacles in complex scenes.(1)TPC-SSD is deployed on the terminal equipment carried by AGV,which is simple to implement and low in cost.Firstly,the network utilizes a learnable downsampling module for foreground points to accurately sample front scenic spots from the original point cloud.Secondly,the voxelated point cloud space is separated and adaptively fused with feature maps of different abstract scales through a multi abstract scale feature extraction module.Then,the center point is predicted from the feature map,and the front scenic spots around the center point are grouped,aggregated,and coded according to different distances using a multi distance scale feature aggregation module to obtain a semantic feature vector.Finally,the center point and semantic feature vectors are used to predict the bounding box.In addition,this article also designs global and local data enhancement schemes for point clouds.(2)VTFC-Net can accurately describe obstacles of various shapes and sizes,making it more complex.It needs to be deployed on a server,and then design a remote communication system for data transmission with AGV.The network first converts a point cloud into a sparse matrix using centroid interpolation voxel operations.In the semantic segmentation branch,the voxel transformer backbone module is used to extract global semantic features,and then predict the classification of each point.In the instance segmentation branch,the 3D sparse U-Net network is used to extract instance features,and then the instance point offset is predicted.Then,the feature completion method is used to predict the instance object grouping for each instance point.Finally,the case integration method is used to integrate excessively segmented instance objects,and the case score prediction method is used to eliminate low quality instance objects.The result of instance and semantic segmentation constitutes the panoramic segmentation result.Through experiments and comparison with mainstream methods,most of the evaluation indicators of the network designed in this thesis reach the first place in official and customized data sets.The network can accurately and efficiently complete obstacle avoidance detection in actual industrial scenarios.