Research On Unmanned Vehicle Optical Target Detection Technology Based On Deep Learning

In recent years,with the rapid development of new sensors and artificial intelligence technology,the use of visible light,laser and other sensors have improved the target detection capability of the unmanned target perception module.However,the existing optical target detection methods based on deep learning cannot meet the high-precision real-time detection of vehicles and pedestrians under the condition of limited on-board computing and storage resources.Therefore,there is an urgent need to design an accurate and efficient unmanned target detection algorithm that can be applied to scenarios with resource constraints.This paper mainly studies the optical target detection technology of unmanned driving scene based on deep learning.The main work and research content are as follows.(1)Aiming at the problems of low background complexity and insufficient sample diversity of the obtained optical target images of unmanned driving scenes,this paper studies and implements a vehicle target data enhancement method based on style coding generation confrontation network.First,this article constructs a generative confrontation network framework consisting of four modules: mapping network,style encoder,generator,and discriminator;Secondly,in order to save the amount of network parameters and calculations,the lightweight structure Ghost Module is introduced as the basic network of the style encoder to generate style codes to achieve high-fidelity vehicle target scene migration,thereby establishing a high-quality extended data set.It provides more abundant data support for the follow-up target detection.(2)Aiming at the problems of low accuracy and slow speed in the current detection of vehicles and pedestrians in the field of unmanned driving,this paper studies and implements an optical target detection method based on sample matching in unmanned driving scenes.The proposed algorithm designs and implements a new type of end-to-end one-stage network based on the combination of residual error and feature pyramid network;In order to improve the detection accuracy of multi-scale targets such as vehicles and pedestrians,dense connections are introduced in the residual module to enhance the ability of feature expression;Finally,in order to balance the high computational efficiency and low storage capacity of the algorithm,a bottleneck structure is introduced in the classification and regression modules,thereby reducing the amount of overall parameters and calculations,and realizing high-real-time and high-accuracy optical target detection in unmanned driving scenes.In order to increase the richness and diversity of the overall data set,random erasure and image blending techniques are used to expand the image.Experimental verification of algorithms on public data sets and self-built data sets,and compared with other typical detection algorithms.The experimental results show that,The algorithm proposed in this paper achieved better detection results.(3)In order to verify the effectiveness of the method proposed in this paper in the detection of unmanned vehicles and pedestrians,this paper designed and implemented an optical target detection software for unmanned vehicles.The software used public and self-built unmanned vehicle image data for offline training,and deployed the trained model on a mobile embedded platform.At the same time,it carried out field experiments and performed online testing and algorithm verification of software functions.The results proved The effectiveness of the method proposed in this paper.