An improved street scene segmentation method was proposed to address the challenges of inaccurate small-object segmentation and high computational cost in autonomous driving. The method was developed based on the DeepLabV3+ framework with integrated attention mechanisms and multi-scale feature extraction. Firstly, the original Xception backbone was replaced by a pre-trained MobileNetV3 network to reduce computational complexity. Secondly, SPM (strip pooling module) was incorporated into ASPP (atrous spatial pyramid pooling) structure to capture elongated spatial features. To address the limitation of single-scale object detection in strip pooling, MSPM (multi-scale strip pooling module) was designed. Finally, deep and shallow features were fused using CGA (content-guided attention) mechanism to preserve essential features, and the Focal Loss function was employed to alleviate class imbalance in the dataset.Experimental results on the Cityscapes dataset indicate that the proposed model achieves mIoU (mean intersection over union) of 59.9% and mPA (mean pixel accuracy) of 90.9%. These represent improvements of 9.73% in mIoU and 0.3% in mPA over the original DeepLabV3+ model. Moreover, the parameter count is reduced by 21.06×10⁶. Comparative evaluations demonstrate that the proposed approach outperforms mainstream semantic segmentation models, including FCN, SegNet, and PSPNet. The study concludes that the method effectively achieves model lightweighting while improving segmentation accuracy for urban street scenes.