Semantic Segmentation Method And Model Implementation For Ground Unmanned Driving

Unmanned driving is one of the important applications of artificial intelligence system in people’s lives,it has the great development prospects.Image semantic segmentation,as one of the important technologies in unmanned driving system,supports unmanned system to analyze the road trafficability.Currently,most models extract semantic feature with the same receptive field.It will cause the model to lose a lot of scale information and reduce the segmentation efficiency of the model.In addition,most models often have a large amount of parameters in order to improve the accuracies.It not only leads to slow calculation,but also increases the demand for hardware resources,which is disadvantage of models to use in unmanned system.This paper mainly studies a lightweight image semantic segmentation algorithm with high accuracy based on convolutional neural networks(CNN),and realizes the application of the method in ground unmanned system.The main research contents are as follows:(1)A lightweight and high accuracy convolutional neural networks for image semantic segmentation is proposed.Low-parameter models can segment input image semantics faster than high-parameter models,and have well real-time performance.Obtaining more multiscale information can improve the segmentation accuracy of the model.Therefore,this paper applies the proposed layer-wise group stacking convolution to reduce the parameters,and uses adaptive selective convolution kernel to extract multi-scale features to improve the segmentation accuracy at the same image area,in order to implement a high-precision lightweight semantic segmentation algorithm.The experimental results show that this method can greatly reduce the number of model parameters,and can achieve effective segmentation results.(2)An image semantic segmentation network via combining context-related information and high-level feature to decode feature map is proposed,in order to improve lightweight semantic segmentation networks.This paper enhances the model’s ability to recover image detailed information of different categories’ boundary during upsampling,by fusing a variety of global context information.The experimental results show that this method can achieve better effects for semantic segmentation.(3)A fisheye image semantic segmentation method based on spatial detail information is proposed.In this method,the spatial detail information extracted from the shallow layer will be input to the deep network through the skip structure to perform information fusion,which improves the performance of the model.Furthermore,in order to train and test the model,the isometric spherical projection model is used to transform the Cityscapes dataset into a fisheye image dataset.The experimental results show that this method can achieve better segmentation results.(4)A semantic segmentation system based on deep convolutional neural networks for urban street scene images is designed.This system can train the model,segment the semantics of the input images and display the segmentation accuracy of different semantic classes according to the corresponding label images.