Study On Traffic Scene Understanding Method Based On Deep Learning

The automobile industry is undergoing major changes.Intellectualization will be the unstoppable development trend of the future automobile industry.The environmental perception based on the car camera is a key technology for realizing the automobile intellectualization.Therefore,it is of great value to study a real-time traffic scene understanding method.Based on deep learning,this paper researches traffic scene understanding methods and proposes a real-time traffic scene understanding method,which lays a theoretical foundation for the practical application of intelligent vehicles.First,based on the analysis of the traditional convolutional neural network model,a semantic segmentation network for traffic scene images is proposed.The network contains an encoder and a decoder.The encoder is designed based on the Alex Net network and is used to extract image features which is used as the input to the decoder.The decoder continuously expands the sizes of the feature maps through the combination of convolution and up-sampling.The size of finally output feature map is the same as the input image size.In order to improve the accuracy of semantic segmentation,the RGB-D image in which the disparity map D and the color RGB image are fused is used as the input of the network.The experimental results show that the proposed semantic segmentation network has good real-time performance.The introduction of disparity map D improves semantic segmentation accuracy to some extent,but the segmentation results are still rough.Then,based on the analysis of the deep residual network and multi-scale network,a multi-scale semantic segmentation network is proposed for the problem of poor segmentation results and poor real-time performance for large-resolution image segmentation.The network structure with multiple scales is designed by merging the up-sampled feature maps with the corresponding feature maps before down-sampling in the shallow part of the network.The experimental results show that the designed multi-scale network has high semantic segmentation accuracy and good real-time performance for the large-resolution image segmentation.Finally,this paper designs a multi-task network to achieve real-time traffic scene semantic segmentation and object detection tasks.The network consists of an encoder,a detection decoder and a segmentation decoder.The encoder structure is established based on the multi-scale network designed in this paper to extract image features which are taken as the input to the segmentation decoder and detection decoder.In the training phase,the detection decoder and the segmentation decoder are separately trained and results are compared with the joint training results.The experimental results show that the proposed multitasking network can achieve high accuracy and good real-time performance in semantic segmentation and object detection tasks of traffic scene images.