面向复杂交通场景的改进DeepLabV3+语义分割算法研究

Research on the Improved DeepLabV3+ Semantic Segmentation Algorithm for Complex Traffic Scenarios

摘要: 针对现有语义分割算法在处理多目标图像时难度大、精度低的问题，改进DeepLabV3+网络提出一种面向复杂交通场景的语义分割算法。该算法改用轻量级骨干特征提取网络，提高计算速度；重构密集空洞空间金字塔模块并引入卷积注意力机制，加强高级特征信息提取；采用多尺度特征融合策略，提高特征恢复精度。在Cityscapes数据集上的验证结果表明，该算法能在保证时效性的同时，准确实现车辆、行人等交通目标的特征提取，为复杂交通环境下的实时分割任务提供可行方案。

Abstract: Addressing the challenges of difficulty and low accuracy in handling multi-objective images with existing semantic segmentation algorithms, an improved DeepLabV3+ network is proposed for a semantic segmentation algorithm tailored to complex traffic scenarios. This algorithm adopts a lightweight backbone feature extraction network to enhance computational speed; reconstructs the dense dilated spatial pyramid module and introduces a convolutional attention mechanism to strengthen the extraction of high-level feature information; and employs a multi-scale feature fusion strategy to improve feature restoration accuracy. The validation results on the Cityscapes dataset demonstrate that this algorithm can accurately extract features of traffic objects such as vehicles and pedestrians while ensuring timeliness, providing a feasible solution for real-time segmentation tasks in complex traffic environments.