Semantic Segmentation of Traffic Scene Based on DeepLabv3+ and Attention Mechanism
Yuan Zhang, Yuhao Zhang, Qianyi Zhang
Abstract
In the increasingly mature autonomous driving technology, acquiring traffic scene information is a guarantee to improve driving safety. The emergence of semantic segmentation technology enables the segmentation of images at the pixel level into regions associated with semantic categories of scenes. This technology improves the vehicle perception and analysis of real-time traffic environment information, thus enhancing the safety of autonomous driving technology. The currently accepted semantic segmentation model DeepLabv3+ still suffers from small targets being ignored and similarly shaped objects being incorrectly segmented in the segmentation task, resulting in a need for further improvement in the accuracy of the result presentation. To address this problem, this paper proposes an improved DeepLabv3+ method for semantic segmentation of road scenes with a fused spatial attention mechanism based on the structure of DeepLabv3+ network, combined with the attention mechanism to increase the weights of segmented regions. After determining the use of Focal Loss, the ASPP structure is improved. The segmentation effects of the models with different ways of introducing the attention mechanism are compared and tested on the Cityscapes dataset. The experimental results show that the improved model segmentation accuracy is improved by 1.56% on average compared to DeepLabv3+, which is an improvement in cross-comparison.