Study And Application On Semantic Segmentation Algorithm With Deep Learning For Road Scene

In recent years,the rapid development of deep learning technology in computer vision has provided support for semantic segmentation applied to road scene recognition of autonomous vehicles.Due to the complex and varied road scenes,the current semantic segmentation scene model has the problems of relatively low segmentation precision,large model parameters and slow speed for complex road scenes,which is difficult to be used in actual terminal equipment.Lane in road scenes are frequently damaged and covered and that does not provide an extension of the lane line.Therefore,this thesis focuses on improving the segmentation accuracy of the model,reducing the number of model parameters and improving the image segmentation speed.Based on this,image semantic segmentation models based on deep learning are designed.The lane in the driving scene is detected separately.In order to obtain higher segmentation accuracy,this thesis improves and experimentally implements the network based on low network from the aspects of short-cut,multi-layer feature fusion,inception model,Max-Pooling and Batch Normalization.These method effectively improve the accuracy of the network,and provide an optimization method for us to build a deeper network.Aiming at the higher segmentation precision in road segmentation,a semantic segmentation model called the first model in this thesis with higher precision is proposed.The model is downsampling based on the residual network,and then the features of different scales are used to obtain features of more scales.These features are output through dense connections to obtain features with multi-scale information,and then obtain the output image by two upsampling steps that the size is same as the input.In order to meet the real-time requirements of automatic driving,a more lightweight semantic segmentation model is obtained.Based on the first model of this thesis,the residual unit is modified,the depth separable convolution is used to reduce the number of model parameters and a complementary model is designed.It can meet the real-time and higher precision requirements of the road semantic segmentation.According to the lane is the important driving track during the driving process and the road semantic segmentation cannot identify the extension direction of the lane line very well in the real world,the two branches are proposed to detection the lines during driving by using the semantic segmentation method without mutual interference.Faster network is used as the basic network to achieve real-time.The two branches achieve semantic segmentation and instance semantic segmentation.In this thesis,each model of the design is trained and tested on the Camvid dataset,Cityscapes dataset and self-made dataset.Finally we get the performance of the model.