Scene Segmentation Algorithm Based On Context Information And Hierarchy

Scene segmentation has broad application prospects in intelligent driving.At present,the scene segmentation algorithm for the intelligent driving mainly relies on semantic segmentation,which aims to recognize and understand objects in the scene.Taking the categories of target objects into account,the scene segmentation can be divided into two categories: scene segmentation based on all objects and human-centric scene segmentation.Extracting the discriminative context features is the crucial component of designing a scene segmentation algorithm.To solve this problem,this work first proposes a novel semantic segmentation model based on dense pyramid module and collaborative learning.The local context pyramid module and the global context pyramid module are arranged in cascade to form the dense pyramid module,enabling extracting the local and global context information simultaneously.As a result,the ability to segment targets of the model is enhanced,in which the recovering of object details and edges,the segmentation of small objects,and the differentiation of different objects get more accurate.In addition,a collaborative learning strategy based on semantic segmentation and object classification is designed to further enhance the ability of the model to extract context information.This approach yields 82.0% mIoU on the Citysacapes test set,demonstrating the effectiveness of the proposed method.Based on the effective idea above,this work further explores the more challenging scene segmentation algorithm based on human beings.Human beings are the most important target in the field of intelligent driving.The fine-grained segmentation of the human body benefits the research of scene segmentation algorithms.However,most of the current algorithms ignore the strong structure within the human body,which limits the development of human fine-grained segmentation.To this end,we divide the human body into three levels,proposing direct inference,top-to-bottom inference,and bottom-up inference to predict the human body at each level,thereby enhancing the flow between the human hierarchy information.Then we compute the confidence of the information source corresponding to the final prediction of each level,and propose the condition information fusion to fuse different information sources,and then generate the final prediction result.The proposed method achieves57.74% and 70.76% of mIoU on the LIP and PASCAL-Person-Part datasets,respectively,outperforming all the previous method,indicating the superiority of the proposed method.