Abstract: In autonomous driving systems, semantic segmentation is a crucial step for perceiving key road elements. Consistency regularization methods, as a representative semi-supervised semantic segmentation approach, can effectively mitigate the reliance of fully supervised semantic segmentation on large-scale high-cost annotated data. Traditional consistency regularization methods inadequately utilize the distribution characteristics of unlabeled data during parameter transfer, thereby constraining their generalization performance. To address these limitations, an adaptive collaborative optimization-oriented semi-supervised semantic segmentation method is proposed. Firstly, a stochastic parameter recovery mechanism that dynamically injects pre-trained model parameters to alleviate noise-induced error accumulation and enhance model robustness is designed. Secondly, a Jensen-Shannon (JS) divergence-based dynamic exponential moving average (EMA) strategy is proposed, which adaptively regulates parameter update intensity by quantifying prediction distribution discrepancies between teacher and student models, effectively addressing weight coupling issues. Finally, an intersection over union (IoU) consistency-based adaptive symmetric cross-entropy loss function is constructed, integrating spatial alignment constraints with bidirectional probability distribution metrics to alleviate overfitting in student model parameters. Experimental results demonstrate that on the Cityscapes and Pascal VOC datasets, using ResNet-50 as the backbone network with only 1/16 of labeled data, our method achieves mean Intersection over Union(mIoU) improvements of 1.09% and 0.86% respectively compared to the state-of-the-art UniMatch method.