Research On Road Scene Segmentation Model Based On FCN And Conditional Random Field

The Advanced Driving Assistant System(ADAS)is the key part of intelligent driving technology and has received much attention and research in recent years.ADAS's main work is to use the vehicle sensor to receive external signals,combined with navigator map data,the system operation and analysis;which access to road traffic information and analysis is the key to the technology.The traditional method is based on image processing,and the algorithm depends on the selection of features.Manually picking a feature rests with the "technique",and it depends heavily on experience and luck whether or not to choose the right feature,which is laborious,timeconsuming.Machine learning rests with the "amount",and the program itself learns the experience and regularity(characteristics and algorithms)from a large number of historical data.The development of deep learning provides a new idea for image processing.Convolution neural network plays an important role in image recognition,detection,segmentation and so on.The main task of road inspection is to rebuild the lane line.If the traffic mark,pedestrian and vehicle target are taken into account,the detection amount involved is too large.On the basis of previous studies,combined with the full convolution neural network(FCN)and conditional random field,this paper discusses semantic segmentation of road scenes,thus replacing multi-target detection.The main work is as follows:(1)This paper presents an improved FCN-based semantic segmentation model.There are two improvements,one is the structural adjustment of the convolution network and the setting of the relevant parameters,and the other is the improvement of the feature fusion method.The main method of deep learning for semantic segmentation is to use the fullly convolutional neural network(FCN)to classify the pixels and reconstruct the segmentation graphs.This paper analyzes the shortcomings of the classical network in the road scene segmentation,and puts forward a new classification network structure,starting from VGG16,cutting the size of the convolution kernel and down sampling to reduce the number of convolutions.In the experimental,the training accuracy of 89%.In addition,based on the two characteristics of FCN8 s,this paper increases the fusion point of pool2,and finally forms the FCN4 s network model,and completes the three-time feature fusion and quadratic upsampling.The experiment compares the semantic segmentation effect of FCN32 s,FCN16s,FCN8 s,and FCN4 s.After quantitative assessment,it found that FCN4 s do have better performance.(2)The algorithm of the conditional random field is integrated into the semantic segmentation model.After the FCN training,the segmentation effect is not bad.However,because of a lack of a priori information,the convolution network has no space,edge constraints,resulting in the slightly rough results.So conditional random field is used in this article for fine tuning.The deep network is responsible for feature extraction,and the probability graph model is used for image refinement.In the experiment,an end-to-end scheme combining FCN and fully connected random field(CRF)is adopted,and the results show that the fusion condition of the random field algorithm does improve the original model of the edge of the details of the training is also more efficient.