Research On Autonomous Navigation Of AGV Based On Machine Vision For Line Following In Complex Environment

AGV(Automated Guided Vehicle)is a multi-purpose robot.It is widely used in various fields such as manufacturing,logistics,warehousing and transportation,and life services.In recent years,with the development of artificial intelligence,AGV has gradually turned intelligent and has become a key tool for intelligent production and services.However,limitations exist in the current navigation of AGV.In the complex environment,the simultaneous fulfillment of navigation accuracy and flexibility is difficult to be achieved by AGV.Although navigation accuracy can be improved through line-following,the navigation still lacks flexibility.Autonomous navigation methods based on laser SLAM,tend to sacrifice navigation accuracy but offer greater flexibility when compared to line-following navigation.Under complex lighting environments such as overexposure,the lane detection based on deep learning is still unable to extract the lane trajectory completely,thus limiting the accuracy of navigation.Additionally,the current lane segmentation network is unable to balance the segmentation speed and segmentation accuracy well,and fails to meet the real-time requirements of embedded computing devices.This paper focuses on the research of AGV navigation in complex environments,bringing improvements and solutions to a series of problems in the current AGV navigation process.The main job of this thesis is centered around:(1)To address the challenge of balancing navigation accuracy and flexibility in AGV navigation,this paper presents a novel approach that combines visual line-following navigation using deep learning with autonomous positioning navigation based on SLAM.In the event of obstacles on the lane,the system seamlessly switches to SLAM autonomous positioning and navigation to enhance the AGV’s flexibility.Moreover,an accurate parking module is incorporated,which enables the calculation of the AGV’s current position information accurately,allowing the AGV to dock more precisely at the designated position.(2)In order to improve the global path planning speed of AGVs during navigation,an optimized A* path planning algorithm is proposed in this paper.The algorithm can dynamically adjust the searching strategy during the searching process to achieve more effective path planning.The addition of the adaptive weight function improves the speed of global path planning.At the same time,high performance is still maintained in terms of path planning accuracy.(3)Aiming at incomplete lane segmentation in overexposure environment,a lane detection method combining image inpainting and segmentation is proposed.In this method,the overexposed lane image is first repaired and reconstructed by the MAE network,and then the reconstructed image is fed into the image segmentation network for lane segmentation.The network parameters of MAE image inpainting model are optimized and reduced,and verified by experiments.The experimental results demonstrate that the inference speed of the MAE model is improved while ensuring the quality of image restoration.A Convolutional Skip Triple Attention(CSTA)image segmentation network is proposed.It meets the requirements of segmentation accuracy and speed at the same time.Meanwhile,the proposed method has better noise suppression.The effect and quality of the segmentation of the network are significantly improved.Finally,the efficiency of the proposed lane extraction method is verified by three image segmentation evaluation metrics(Io U,F1-score and PA)and inference time in the case of overexposure and proper exposure.(4)The overall navigation control decision node is designed.The control data of ROS basic navigation,visual patrol navigation and precise parking module are fused.The AGV is made to switch to different navigation modes in varied situations.At the conclusion of this paper,all the functions of visual line inspection,SLAM navigation and obstacle avoidance,and precise parking are tested through experiments.The experiments demonstrate that the proposed navigation method can maintain high accuracy even in overexposed environments.Furthermore,the utilization of multiple navigation modes ensures that the AGV navigation system remains flexible,thus increasing the accuracy and adaptability of AGVs in complex environments.