| In recent years,indoor mobile robots have diversified their application scenarios under the dual impetus of application demand and market economy,such as "robot+" application action,market intelligence transformation,and have put forward higher requirements for reliable operations.However,indoor mobile robots still have problems such as poor environment perception ability and single detection strategy,which leads to low reliability of obstacle detection and difficulty in ensuring the safety of robot operation,which hinders the large-scale application and promotion of robots.This paper studies the obstacle detection technology of mobile robots based on laser point cloud and visual semantics fusion,which has important practical application value to promote the application deployment and intelligent transformation of robots in complex scenes.The main research contents of this paper are as follows:In order to meet the application requirements such as strong perception capability,high reliability and good real-time performance in complex scenes of mobile robots,the overall system of mobile robot platform is designed from both software and hardware systems,including mechanical structure ontology design,sensor selection and software system architecture construction;the differential odometer,laser odometer and environment map model are established according to LIDAR,monocular camera observation model and coordinate transformation,and the overall technical framework of camera distortion correction and obstacle detection is completed.Aiming at the sensor data instability caused by LIDAR installation error and drive wheel size wear,the online calibration method with internal and external parameters based on trajectory matching is proposed.The experimental results show that the relative error of distance parameter is ≤2.5mm,the relative error of angle parameter is ≤0.2°,and the trajectory matching error is reduced by 1.1 times.Aiming at the interference points such as original point cloud noise and dispersion of Lidar,the effective target point cloud is extracted by preprocessing.Aiming at the laser point clouds connectivity problem,the ground point cloud segmentation algorithm based on adaptive slope threshold is proposed.The experimental results show that the accuracy and recall rate of this method are>0.850,the average time is 22.7ms,and the comprehensive performance is better than other methods.Using the graph-optimized laser SLAM algorithm to build an 18.2m*16.5m indoor environmental raster map with relative error of linear features ≤0.83%and relative error of angular features ≤1.446%,which can provide accurate environmental a priori information for obstacle detection.Aiming at the problem of single obstacle detection strategy,the panoramic adaptive planar obstacle detection algorithm based on priori maps is proposed.The experimental results show that the proposed algorithm can effectively solve the problems such as escape and security risks in narrow and high-speed scenes,and the processing time of single frame data is≤18.64ms.Aiming at unreliability of data fusion between monocular camera and Lidar,online external reference calibration is completed by using graph-based optimization idea.YOLOv5 network architecture and transfer learning scheme are used to complete obstacle semantic model training of self-made data set.Aiming at a single plane information cannot reflect the semantic and spatial dimension information,the obstacle detection algorithm based on visual semantics and distance correlation is proposed.The experimental results show that the average positive detection rate of the proposed method is 0.915,which is better than the traditional method.The error of length and width is ≤10cm,and the detection time is ≤37.7ms.Aiming at different information fusion of obstacles,the multidimensional obstacle information fusion strategy based on plane mapping is proposed.The experimental results show that the robot can accurately perceive the spatial dimension information of obstacles and realize dimensional expansion processing according to semantic information when it passes through hollowing and dynamic scenes,and the width relative error<10cm,which can effectively improve the reliability of obstacle detection. |
PDF Full Text Request