Research On Real-time Semantic Segmentation Algorithm For Urban Road Scene

With the advent of the era of artificial intelligence,a variety of intelligent service devices gradually appear in life,such as intelligent medical,intelligent home and unmanned aerial vehicles,and automatic driving occupies an important position among them.Therefore,improving the level of semantic segmentation technology applied to autonomous driving,enabling it to achieve excellent performance in three aspects: segmentation accuracy,reasoning speed,and parameter scale,enabling intelligent vehicles to accurately and timely analyze and grasp road conditions,is an important factor in ensuring safe and smooth autonomous driving.Therefore,this paper conducts research on real-time semantic segmentation algorithms for urban road scenes,with the main contents as follows:First,a Lightweight Asymmetric Network based on Attention and Dialated convolution using Encoder-Decoder structure,namely AD-LANet,is proposed.It is mainly composed of Improved Deep-wise Asymmetric bottleneck Residual module(IDAR)and improved attention module(ISA and ICA).The IDAR module is based on a deep asymmetric bottleneck module,combining deep separable convolution,dialated convolution,and feature interaction operations;ICA module and ISA module introduce pyramid pooling structure and void convolution respectively.Secondly,based on AD-LANet,an Efficient Lightweight Asymmetric network based on Attention and Dialated convolution,namely AD-ELANet,is designed.In the encoder part,the channel shuffling operation is injected into the IDAR module,and an Improved Deep-wise Asymmetric bottleneck Residual module with Shuffling operations(S-IDAR)is proposed,which can make the information between different channels more effectively blend with each other;In the decoder part,the ISA module and ICA module are further modified,and the improved attention module-second generation(ISA-2 and ICA-2)is proposed,and the Improved Feature Fusion module(IFF)is built to process the feature information more efficiently.Finally,on the Cityscapes dataset and the Cam Vid dataset for urban road scenes,the two real-time semantic segmentation models were designed for ablation and comparison experiments,and verified and evaluated.It can be seen from the experimental results that both AD-LANet and AD-ELANet have achieved a good trade-off between segmentation accuracy and inference speed,and achieved a good segmentation effect with fewer parameters.