Research On Anchor-free Instance Segmentation Algorithm Of City Street Scenes

Due to the rapid development of deep learning technology and computer vision,the intelligence of autonomous driving has been greatly improved,and low-cost images are used to analyze complex scenes.Instance segmentation is not only a simple detection of objects or clustering pixels of the same category,but is a combination of the two tasks,which can better establish an understanding of the scene and become the most challenging task in computer vision.The instance segmentation algorithm based on deep learning of city street scenes is studied,and the main contents of the work are as follows:First of all,this thesis proposes the adaptive context network for instance segmentation.The weighted FPN is designed to combine the position information of the shallow feature maps with the semantic information of the deep feature maps by feature weighting.The channel attention and spatial attention modules are proposed and applied to the regression branch to better integrate spatial information and channel information.The adaptive context-guided mask branch is designed,and the segmentation mask is predicted pixel by pixel by fusing global features and local features of different regions.Secondly,this thesis proposes an instance segmentation algorithm based on dynamic filters,which is composed of an improved object detector and a mask branch based on dynamic filters.Apply the improved receptive field block to the regression branch to integrate the feature maps of different receptive fields.The attention-based dynamic filters are designed and applied to the mask branch to improve the quality of the prediction masks,which can better integrate the features through the position attention module and the channel attention module.In order to control the amounts of parameters and speed up the inference time of the model,the dynamic filters perform convolution on the mask branch by the method of depth separable convolution.In this thesis,verify the rationality of the algorithm by a large number of ablation experiments on the city street scenes dataset Cityscapes,and verify the advanced nature of the algorithm by comparing them with a variety of the latest instance segmentation algorithms.The comparative experiment is done on the COCO dataset to verify the effectiveness of the algorithms.The trained models are tested in domestic city street scenes respectively,and the comparative analysis is made on the visualization renderings to verify the universality of the two models.Experiments show that two algorithms,which are the adaptive context network for instance segmentation and the dynamic filters for instance segmentation,are better than the existing state-of-the-arts to deal with city street scenes,and the latter has a better segmentation effect on city street scenes.