Research On Object Detection Method For Unmanned Vehicle

Unmanned vehicle is a vehicle that can operate autonomously without a human driver.The target detection is an important task in the perception and localization module of unmanned vehicle systems.The accuracy and robustness of the target detection algorithm will greatly affect the rationality of the instructions issued by the decision-making and planning module.Based on the path planning and obstacle avoidance requirements for unmanned vehicle,this paper focuses on researching lane line detection methods and obstacle detection methods for unmanned vehicle.1.Aiming at the characteristics of diverse lane shapes and strong external interference when the unmanned vehicle is running,a lane line detection method based on improved YOLOP is proposed,which is composed of feature extraction module,feature fusion module and segmentation head module.The feature extraction module extracts feature maps of different scales,while the feature fusion module uses a feature pyramid network.In the first layer of the feature fusion module,a Simplified Spatial Pyramid Pooling-Fast(Sim SPPF)module is added to expand the receptive field and merge feature maps of different scales.To address the elongated geometric features of the lane lines,two Coordinate Attention(CA)modules are added between the feature extraction and feature fusion modules.By assigning different weights to the feature maps,more important features receive higher weights during parameter updates,allowing for better feature selection.The segmentation head module restores the feature maps to the input size and outputs predicted mask results,representing the probability of each pixel belonging to a lane line.Experiments conducted using BDD100 K dataset validated the effectiveness of the proposed lane detection method.2.Aiming at the characteristics of multi-source and heterogeneous detection data and the difficulty of feature learning in obstacle detection problems,a obstacles detection method based on Frustum-Linked Dynamic Graph CNN(FrustumLDGCNN).This method consists of three modules: frustum proposal generation,point cloud feature extraction,and 3D bounding box prediction.The frustum proposal generation module uses a 2D detection method to generate 2D boxes on the image,then projects the regions where the 2D boxes are located onto a 3D point cloud,and transforms the coordinate system to a unified coordinate system to form a view frustum point cloud proposal region.This fusion method can avoid the calculation of a large number of background points and to some extent avoid the interference of noise points,thereby improving the accuracy of detection.The point cloud feature extraction module uses Linked Dynamic Graph CNN(LDGCNN)network to extract point cloud features in the view frustum candidate region,which can learn features between points and provide rich point cloud spatial feature information and output the probability scores of the target points.Finally,the 3D bounding box prediction module is used to locate the target object and predict its 3D bounding box.Experiments with the KITTI dataset have verified the effectiveness of the obstacle detection method proposed in this paper.A target detection prototype system for unmanned vehicle driving has been designed and developed.The improved YOLOP lane detection method and FrustumLDGCNN detection method are integrated.The client interface and system functions are designed from the user’s perspective.