Environment Perception On Field Road Based On Convolutional Neural Network

At present,the driverless car technology in urban road environment has received more attention and achieved many outstanding achievements,but there are still many chalenges in the driverless car technology in the field road environment.How to realize efficient and accurate perception of vehicle environment in the complex and changeable field road environment is one of the key technologies to realize the field unmanned driving.This paper focuses on the use of convolutional neural network to realize the autonomous perception of unmanned mobile platform in the field road environment.The research work of this paper is mainly divided into three parts: the field road environment perception technology based on the classification network,the field road environment perception technology based on the semantic segmentation network and the vehicle experiment based on the convolutional neural network.Main contents completed are:In this paper,the field road environment perception without pedestrians is simplified as a classification problem,and a classification network is used to make a rough perception of the field road environment without pedestrians.This paper improves the backbone network in ShuffleNetV2 network and proposes a lightweight classification network of FSNet.The accuracy of FSNet network is 97.28%,and the processing time of single image is 5.94 ms,which is 49.5% higher than ShuffleNetV2 network.The task of sensing the field road environment with pedestrians is relatively complex.This paper uses semantic segmentation network to segment the road environment in front of the unmanned mobile platform at pixel level.The semantic segmentation network has poor generalization and is prone to misidentification in the field road environment perception task.The DGNet semantic segmentation network proposed in this paper can solve these problems wel.The DGNet network is built on the basis of DeepLabV3 network.The grouped atrous spatial pyramid pooling is used in multiple network layers.The up-sampling part of the network uses the DupSampling method.DGNet network has better network performance than DeepLabV3 network.The MIoU of DGNet network is 0.856 and the single image processing time is only 15 ms.At the same time,this paper collects the images of the field road environment by itself,and makes a large-scale semantic segmentation datasets.In this paper,the FSNet network and DGNet network are deployed on the unmanned mobile platform to sense the environment of the road ahead of the unmanned mobile platform.At the same time,the controller based on FSNet network and DGNet network is constructed to guide the autonomous driving and dynamic obstacle avoidance in the straight-running stage of the unmanned mobile platform.Experiments show that the perception system based on the convolutional neural network can quickly and accurately perceive different field road environments,and realize autonomous driving and dynamic obstacle avoidance in the straight running stage of the unmanned mobile platform with very little human intervention.This paper contains 47 pictures,12 tables and 65 references.