Research On Visible And Infrared Images Semantic Segmentation For Autonomous Vehicles

The automatic driving system mainly includes the perception system and the decision system,among which the semantic segmentation is one of the important technologies in the perception system.At present,semantic segmentation algorithms based on visible images have made great progress.However,it is known that in environmental conditions with low light,glare and fog,the detection performance of visible light cameras is limited,which affects the accuracy of the perception system finally.Considering the advantages of infrared cameras in the above environment,a semantic segmentation algorithm of visible light and infrared fusion is proposed in this thesis.The algorithm makes full use of visible light and infrared image information and can be used in automatic driving perception.In this thesis,the deep learning method is adopted to design the semantic segmentation algorithm.The main work includes the improvement of the semantic segmentation data set and neural network and the optimization and acceleration of the semantic segmentation model.The main contribution is summarized below.First,the original multi-spectral data set had some shortcomings.The distribution of the data set was modified more reasonably for the first time,and the improved multi-spectral data set was made by data augmentation using basic image processing methods.Second,based on the existing classical and fusion semantic segmentation neural networks,it is proved that the infrared image information is beneficial to improve the image segmentation accuracy in night scenes.It is found that the double-branch encoder in the fusion network has no advantages over the single-branch encoder in improving the network performance.Third,in order to solve real automatic driving environment scenes,from the perspective of improving accuracy and speed of the algorithm,this thesis modified the loss function of the neural network and use quantization,pruning,efficient convolution and other acceleration technologies.The proposed semantic segmentation algorithm for visible and infrared images takes into account both accuracy and speed,which has important value for the development of autonomous driving perception system technology.