Research On Unstructured Road Environment Perception Technology Of Mobile Robot

In recent years,with the rapid development of modern high technology,the research on autonomous navigation of mobile robots has been intensified,and the environment perception technology is the foundation of autonomous navigation for mobile robots.At present,the environment perception technology in structured road environments such as cities and highways has received more attention and achieved many excellent results,but there are still more challenges in environment perception technology in unstructured road environments in the field.Unstructured road environments do not have regular road boundary lines,specific colors and features,and clear edges.To address the complex and diverse unstructured road environment,this paper adopts convolutional neural network technology to study the unstructured road environment perception technology,constructs an environment perception system,realizes the accurate and efficient autonomous perception of the surrounding environment by the mobile robot,and realizes the autonomous driving and dynamic obstacle avoidance of the mobile robot based on the results of environment perception.The main research contents of this paper are as follows.(1)Research on the detection method of drivable area for unstructured road environment.In view of the characteristics of unstructured road environment,a semantic segmentation method is used to divide the driveable road area in front of the mobile robot,and a lightweight DSSNet semantic segmentation network model is constructed for the mobile robot application platform with poor hardware computing capability.activation function instead of Re LU activation function,while improving the ASPP module in Deep Lab v3,using the grouping method to extract features from multiple layers,and using the DUp Sampling upsampling method for the inverse coding process.The semantic segmentation dataset is collected and produced,and the performance of DSSNet semantic segmentation network is tested on the self-harvested dataset.The semantic segmentation network proposed in this paper can reduce the segmentation error of the network and improve the operation speed of the network.(2)Research on pedestrian detection method based on acceleration optimization.In order to realize the dynamic obstacle avoidance function of mobile robot in unstructured road environment,YOLOv3 algorithm is used to detect pedestrians accurately.The YOLOv3 algorithm is studied in detail in terms of network structure,bounding box prediction principle and loss function,and the dataset for pedestrian detection is constructed and analyzed for YOLOv3 and other mainstream target detection algorithms,and the results show that YOLOv3 algorithm performs better in terms of detection accuracy and detection speed.Finally,in order to meet the demand of running real-time on mobile robot platform,the network model is accelerated and optimized using Tensor RT to meet the requirement of real-time on mobile robot.(3)Research on pedestrian localization method based on vision and LIDAR fusion.The image data of monocular camera can only detect pedestrians and cannot get the location information of pedestrian targets.This paper proposes a method to detect pedestrians by vision and localize them by LIDAR at the same time to obtain complete pedestrian information.The camera and LIDAR are jointly calibrated,and the fusion algorithm is used to achieve temporal and spatial matching between LIDAR and camera.The fusion method of vision and LIDAR is tested,and the localization error of pedestrian targets reaches within 6 cm,which meets the requirements of environment perception of mobile robots.(4)Construction and experiment of unstructured road environment perception system.The unstructured road environment perception system of the mobile robot is constructed,and the road detection results based on the lightweight DSSNet semantic segmentation network model,the accelerated and optimized YOLOv3 pedestrian detection results and the pedestrian localization results of vision and LIDAR fusion are mapped to the control of the mobile robot.The individual modules of environment sensing are deployed on the ROS robot operating system and robot control strategies are developed for both the no-pedestrian and pedestrian cases.Finally,the environment sensing system is experimented in an actual field unstructured road scenario,and the autonomous driving of the mobile robot in the absence of pedestrians and the dynamic obstacle avoidance of the mobile robot in the presence of pedestrians are achieved with minimal human intervention based on the results of environment sensing.The thesis includes 67 figures,20 tables,and 82 references.