Research On Semantic Segmentation For Autonomous Drivin

With the rapid development of artificial intelligence and computer hardware devices,researchers have paid great attention to autonomous driving technology.Scene understanding plays a crucial role in decisionmaking for autonomous driving tasks,primarily referring to the perception and understanding capabilities of the surrounding environment,including the recognition and classification of various elements such as roads,vehicles,pedestrians,and traffic signs,as well as the analysis and judgment abilities for complex driving scenarios.Autonomous driving scene understanding technology typically adopts deep learning semantic segmentation techniques to perceive and process environmental information around the vehicle.This article presents a deep learning-based semantic segmentation algorithm that provides a semantic segmentation method for scene understanding in autonomous driving tasks.The research focuses on the following aspects:(1)This article provides a detailed study of the advantages and disadvantages of various structures,starting from a lightweight design approach,using the branch structure network Bi Se Net V2,and conducting a detailed study of this network(2)To solve problems where ordinary convolution cannot handle large receptive fields or irregular sampling rates in semantic segmentation tasks,this thesis uses dilated convolutions and deformable convolution modules respectively.Dilated convolutions expand receptive fields by introducing holes into convolution kernels to better capture relationships between distant pixels;deformable convolutions can deform convolution kernels to adapt them to irregular objects within images.(3)Pyramid pooling can improve model accuracy by pooling at multiple scales helping models capture features at different scales.This thesis introduces a multi-path fusion pyramid pooling module which combines feature information from different scales together to improve network’s segmenting accuracy.The impact on network performance with different sizes of poolings was verified through extensive experiments.In addition,to avoid overfitting of training data,this article also introduces the Drop Block module to improve network generalization ability.This thesis proposes a lightweight network-based semantic segmentation method for automatic driving from the perspective of lightweight networks.It provides strong support for achieving more accurate and reliable automatic driving.