| Nowadays,various types of robots are emerging,and autonomous navigation robots that can independently perform operations such as environment perception and path planning have gradually become the focus of research.However,the perception system of a single autonomous navigation robot has limitations.When obstacles are far away from the robot,there will be perception blind spots.With the development of wireless communication technology,multiple robots can exchange perception data through wireless communication to expand the perception range,and perception cooperation has gradually become an important solution to improve the driving safety of autonomous navigation robots.There are three types of shared perception data: raw data,feature data,and output-level data.Raw data can preserve the original geometric characteristics of objects to the greatest extent,but the large amount of data cannot guarantee real-time performance;feature data can retain important information of objects,but the requirements for fusion are high and the universality is poor;the data volume of the output-level data is small,the universality is strong,and it is easy to fuse,but it cannot sense the missed obstacles.The above three kinds of data have their own advantages and disadvantages,and Current fusion methods cannot eliminate the influence of sensor errors.Based on this situation,the main research work of this paper is as follows:1.This paper proposes a new type of shared data: a combination of 2.5D geometric map and obstacle list,which not only includes output-level data,but also includes a geometric map that can describe the height characteristics of the surrounding environment,this kind of shared data can provide the height information of missed obstacles,which solves the problem that the output level data cannot perceive missed obstacles.2.This paper proposes a geometric information generation algorithm: in the process of pillar-based object detection,the height features in each pillar area in the original 3D point cloud are extracted to generate a 2.5D map.In order to reduce the transmission time,we use Octomap to convert the 2.5D geometric map into a byte stream representation.This algorithm can be paralleled with the object detection algorithm and does not use the intermediate feature data output by the convolutional neural network.It has strong universality.3.This paper proposes an improved fusion method: The VGICP registration algorithm is added to register the shared data,the algorithm can help reduce the impact caused by the sensor’s error,which can help the robot generate a more accurate global map.Finally,this paper uses ROS Gazebo virtual simulation software to build a virtual city environment for simulation testing,and designs multiple sets of comparative experiments to compare the perception cooperation algorithm for transmitting different types of shared data with the algorithm proposed in this paper.From multiple dimensions such as shared data volume,total system time,and shared data sending frequency,we verify the feasibility and effectiveness of the algorithm.The experimental results show that the algorithm can help detect obstacles that are not classified by the object detection algorithm,and the accuracy of detection can reach85.71%,which is 5.5% higher than the Cooper algorithm that uses sensor raw data as shared data.The size of the proposed shared data is only 33.47 KB.Under 4G communication conditions,the average transmission delay is only 12.68 ms,which is98.42% lower than the Cooper algorithm.Transmission of perception sharing data helps autonomous navigation robots improve detection accuracy and generate more accurate global maps. |
PDF Full Text Request