The Logistics Unmanned Vehicles Targets Detection Based On Deep Learning

In recent years,with the logistics unmanned vehicle distribution service is operated,in order to ensure the safe and reliable driving of logistics unmanned vehicles,the high-precision detection of targets through sensors such as vehicle-mounted lidar and cameras become a major logistics and e-commerce business.issues of general concern to the company.However,the existing 3D target detection methods are generally difficult to extract point cloud features,and the fusion detection model is difficult to fully extract multi-modal information,resulting in the inability to effectively detect road targets,and logistics unmanned vehicles still exist in the realization of distribution tasks.Lots of safety hazards.Based on this,this thesis focuses on the target detection of logistics unmanned vehicles based on deep learning.The main work and research content were as follows.(1)Aiming at the sparseness,imbalance characteristics and spatial irregularity of point cloud,which leads to the target points are scarce and the feature extraction of region proposal is difficult.This thesis studies and implements a A point cloud target detection method based on graph neural network and attention mechanism.Combining the graph neural network with the attention mechanism to extract features from the points of interest can extract more abundant information from region proposal,so that the target can be accurately detected only by relying on a small number of point clouds.At the same time,a point cloud density perception radius prediction module is studied,which can dynamically adjust the feature extraction range of candidate regions according to the different point cloud density distributions,avoid the invalid feature extraction operation when the point cloud data is empty and save the computational cost to a certain extent.It improves the adaptability of the network model to different degree of point cloud sparsity.Compared with other typical point cloud target detection methods,the point cloud target detection network studied in this thesis had been a good performance in detection speed and detection accuracy.(2)Aiming at the problems of low detection accuracy and difficulty in adapting to complex scenes caused by the difficulty of integrating 3D scene context information into the multimodal feature fusion layer in the existing laser point cloud and visible light image fusion detection mechanism,this thesis has studied and implemented an end-to-end feature fusion detection method based on Transformer.Using Transformer’s attention mechanism to integrate 3D scene context information into multi-stage feature fusion layers of different modalities,fully integrate information from different modal data,and overcome the problem of insufficient extraction of different modal information by traditional fusion mechanisms,which effectively improves the target detection accuracy.Through the comparison and analysis of multiple sets of experimental results,the method in this thesis had been better performance in detection accuracy.(3)In order to verify the effectiveness of the target detection method in this thesis,a logistics unmanned vehicle simulation platform is built in this thesis,and a large amount of data has collected by sensors such as lidar and camera mounted on the platform.At the same time,in order to further realize the visualization of point cloud and image,this thesis has designed and implemented the target detection software for logistics unmanned distribution scenarios.Through the application experiments in different outdoor scenarios,the function of the software has been verified,and the goals proposed in this thesis had been proved.The effectiveness of the detection method in the logistics unmanned vehicle delivery scenario.And had been proves the effectiveness of the target detection method in this thesis in the logistics unmanned vehicle distribution scenario.