Free Space Detection Based On Weakly Supervised Semantic Segmentation

With the rapid development of deep learning,semantic segmentation has become feasible and practical in the field of free space detection.Since fully-supervised methods rely on a large number of expensive pixel-level labels,we can utilize more readily available imagelevel class labels for free space detection based on weakly supervised methods.Most existing weakly supervised semantic segmentation(WSSS)methods achieve pixel-level segmentation by associating class labels with spatial information of image to locate the target region.However,autonomous driving scenes are more complex than the nature scenes for which most WSSS methods are applicable.Repeated and more salient nonroad objects will lead to ineffective learning of the discriminative features about free space and bring about wrong spatial visual associations,so the free space detection results still need to be improved.This thesis proposes an effective weakly supervised framework suitable for free space detection under complex autonomous driving scenarios.The main contributions are as follows.(1)In order to achieve effective association between class labels and free space visual-spatial information,a weakly supervised free space detection method based on network attention correction is designed.The network’s attention is focused on the road by adopting the class-balanced sample collection and the bottom-crop strategy,so that the classification network can effectively learn more discriminant features about free space and achieve accurate locating.Then we introduce the multi-scale context extraction module and the equivariant regularization module to propose the Res Net-38-d ER model.Res Net-38-d ER obtains denser pixel-level spatial information and equivariant consistency constraints,which further improves the accuracy of the detection results.(2)In order to optimize the boundary details of the detection results,this thesis develops a weakly supervised free space detection framework by fusing spatial priors and region features.The method utilizes superpixels to provide boundary information,and generates region features with local semantic consistency to characterize appearance similarity of free space.Then,we design an adaptive weighted clustering algorithm SWKmeans to refine the boundary of the detection results by fusing the localization information with the region features.The methods proposed in this thesis are compared with other methods on the autonomous driving dataset Cityscapes.The Io U metrics on Cityscapes validation set and test set reached 0.847 and 0.860,respectively.Extensive experimental comparisons show that our work is an effective and robust method to perform free space detection in complex environments using only image-level annotations.