Research On Segmentation Method For Scene Of Urban Street Based On Deep Learning

With the rapid development of social economy,the number of motor vehicles has exploded,and the difficulty of management has increased,resulting in road congestion and frequent traffic accidents.The intelligent city street view detection system plays an important role in traffic management and effectively improves the level of traffic management.However,due to the interference of many small objects such as vehicles and pedestrians and the objective factors of object occlusion in the urban streetscape environment,the segmentation is more difficult,and false detection and missed detection often occur,which seriously affects the segmentation effect of urban streetscape.Therefore,based on the theory of deep learning,this thesis researches the instance segmentation method of urban street view images.In order to reduce the missed detection and false detection caused by object occlusion,an instance segmentation algorithms for fusion data processing tasks is designed.A data enhancement method based on mask reconstruction is proposed.The urban street view image is masked and reconstructed by a lightweight asymmetric encoder-decoder architecture.The encoder operates in the visible area,and the decoder reconstructs the lost area in the pixel space.The post-processing part uses the super-pixel image of the urban street view data to enhance the predicted instance mask.In the edge area of the instance mask,when the ratio of the area of the super-pixel unmasked area to the masked area is lower than a certain threshold,the super-pixel is added to the instance mask.By comparing the segmentation results of Mixup,Cutout,Random Erasing and the proposed algorithm,the segmentation ability of the instance segmentation algorithm for occluded objects is verified.Aiming at the problem of small object segmentation in urban street scene,the SOLO instance segmentation algorithm is improved by feature weighting and multi-scale feature fusion,and the F-SOLO(Feature Enhancement SOLO)instance segmentation algorithm is proposed.After the feature extraction network,the feature channel is weighted and reconstructed to improve the utilization of important features.The feature fusion module extracts features of different scales,and four convolution kernels with different expansion rates aggregate the features of different channels according to periodic transformation to achieve finer-grained utilization of convolution kernels.The experimental results show that the mAP(mean Average Precision)of F-SOLO algorithm on Cityscapes dataset is 27.2%,which is 2.2% and 2.1% higher than that of single-stage SOLO algorithm and two-stage Mask R-CNN algorithm,respectively.The instance segmentation algorithm is applied to urban streetscape detection,and an intelligent urban streetscape detection system is designed.The system can continuously detect and segment pedestrians and vehicles in the road,and realize intelligent pedestrian safety warning,prohibited vehicles and vehicle violation alarm functions.