Research On Semantic Segmentation Of Unmanned Driving Scenes Based On Limited Annotated Data

Semantic segmentation is one of the basic tasks of environmental perception and understanding of unmanned driving systems.However,the high labeling cost,the systematically mislabeled annotation in complex scenes,the variability of the task environment of unmanned driving systems and the poor generalization ability of segmentation models seriously restrict the application of semantic segmentation in unmanned driving systems.This paper focuses on the theme of reducing the dependence of semantic segmentation algorithms on labeled data and achieving high-precision semantic segmentation of unmanned driving scenes under the condition of limited labeled data.The specific work is as follows:1.In order to solve the problem that the segmentation model is difficult to converge in traditional pretraining and supervised-training paradigm under limited labeled data,this paper proposes a self-supervised learning semantic segmentation algorithm based on multi-scale spatial constraints.By applying the proposed multi-scale dense representation self-supervised learning algorithm based on space constraints on unlabeled data close to the task scene to pretrain the model,the model can produce a good feature response to the task environment in the pretraining stage.Furthermore,a committee voting-based sampling strategy is incorporated in the training process of the segmentation model.Experiments show that the segmentation accuracy of our model exceeds the traditional training paradigm by more than 3.94% under various amounts of labeled data.2.In order to make full use of the limited labeled data and a large number of easily obtained unlabeled data in the task scene,this paper proposes a semi-supervised semantic segmentation algorithm based on label augmentation and adaptive mutual supervision.For labeled data,the sequence frame information is used to propagate the labels to adjacent frames through the video prediction model.At the same time,this paper adopts a new labeling strategy for labeling key regions of the sample.For unlabeled data,this paper uses class-balanced dynamic sampling,anti-focus loss and branch mutual supervision strategy to prevent the homogeneity of model branches while filtering out the consistency regularization noise.Experiments show that our approach achieves the stateof-the-art segmentation accuracy among semi-supervised segmentation methods under various amounts of labeled data.3.Aiming at the situation where no labeled data is available in the task environment,this paper proposes a two-stage unsupervised domain-adaptive semantic segmentation algorithm based on the mutual enhancement of semantic and deepth features.In the first stage,the domain representation capability of the model is enhanced by the mutual enhancement module of semantic and deepth features,then fine-grained cross-domain feature alignment is achieved by constructing depth-aware domain feature and adversarial learning based on pixel-level semantic category recognition.In the second stage,selfcorrection training based on class centroids is performed to make the feature distribution more compact.Experiments show that our approach achieves 46.32% m Io U in Cityscapes validation set,which provides a feasible option for the rapid migration of the semantic segmentation model for unmanned driving systems.