Road Scene Segmentation Based On Fully Convolutional Neural Networks

In recent years,autonomous driving technology has become a focus in many countries.Autonomous driving technology is mainly divided into three parts,one is perception,one is decision-making,and the other is control,and the most basic part is the car's environmental perception module.However,the complexity of road scenes poses great challenges and difficulties to the perception and understanding of the surrounding environment.So the cost of the existing road scene perception system is extremely high.If the road scene perception technology is empowered by computer vision solutions,the cost will be greatly reduced.Semantic segmentation is a very powerful road scene understanding technology,which can perform pixel-level classification and recognition of images,and can well perceive road scenes,and has wide practical values.At present,the road scene semantic segmentation technology based on deep learning has reached a level that fully surpasses the traditional semantic segmentation ones in terms of segmentation accuracy and speed,but there are still many problems,such as:(1)The object is too large resulting in incomplete segmentation;(2)The segmented object is too small to be recognized;(3)The segmentation target is greatly affected by factors such as lighting,which makes the target not obvious.To address the above problems,this thesis proposes a semantic segmentation model based on fully convolutional neural networks.The main contributions are as follows:(1)An encoder-decoder semantic segmentation model is introduced.The encoder part adopts Resnet18 Dilated structure with dilated convolution.Dilated convolution is used to increase the receptive field without reducing the image resolution.(2)Different pooling is applied to obtain the feature maps of different receptive fields,to solve the problems where the segmented objects exceed the receptive field and thus cannot be completely segmented,or the segmented objects are too small to be recognized;(3)A dual attention mechanism is employed to establish a local dependence on the overall features,and thus solve the problem of obscuration of the targets and different scales caused by the occlusion factors such as lighting;Through extensive experiments on the authoritative data sets ADE20 K and Citysacpes,the model proposed in this thesis is compared with several state-of-the-art semantic segmentation models.The experimental results show that the model in this thesis can significantly improve the classification accuracy,and can address the deficits of the current models to a certain extent.