Traffic Image Segmentation Based On Deep Residual Fully Convolution Neural Network

Traffic image segmentation is to identify and segment all participating objects such as vehicles,pedestrians,and road surfaces from current traffic images.It is an important part of smart driving and intelligent transportation systems.Real-time and accurate traffic image segmentation helps improve traffic safety and improve driving efficiency.Traditional traffic image segmentation methods are difficult to meet the requirements of real-time and accuracy at the same time.However,segmentation methods based on deep convolutional neural networks usually have strong characterization capabilities and good application potential.Therefore,it is of great theoretical and practical significance to study traffic image segmentation method based on deep convolutional neural networks.In this article,a traffic image segmentation method based on deep residual fully convolutional neural network was designed and implemented,and its effectiveness was verified on experiments.Firstly,the research on traffic image segmentation benchmark datasets and image preprocessing method was carried out.And then,a deep residual fully convolutional neural network model was designed,which combined the advantages together including that deep residual network have strong feature representation capability,atrous convolution can amplify the receptive field and fully convolutional neural network can achieve pixel-level segmentation.Finally,for the benchmark dataset and the self-collected dataset,the 97.03% and 85.99% MIOU(Mean Intersection Over Union)and the 0.065 s and0.370 s average segmentation speed of per images was obtained.Compared with the conventional fully convolutional neural network,the results show that the segmentation accuracy and real-time performance of the proposed method are improved.The work mainly has the following two parts:1.Research on benchmark datasets and image preprocessing methods.Firstly,the benchmark traffic image segmentation datasets were explored and another domestic traffic image dataset was collected and labeled.Compared with the benchmark dataset,this self-collected dataset included more domestic urban and highway traffic scenarios.And then,as for image preprocessing methods,contrast-limited adaptive histogram equalization,image pyramid and image cropping were used to improve image quality and diversity.2.Research on the suitability of deep learning models and related functional modules.The deep neural network models such as convolutional neural network,fully convolutional neural network and deep residual network were elaborated and the suitability of conditional random field and atrous convolution function modules in traffic image segmentation were also described.Firstly,a segmentation model based on fully convolutional neural network and conditional random field was proposed,which had high precision but no enough real-time performance.Then,a segmentation model based on deep residual fully convolutional neural network was designed and implemented,and the accuracy and real-time performance are improved.