Research on Semantic Segmentation Algorithm for Autonomous Driving Based on Improved DeepLabv3+

The current mainstream high-precision semantic segmentation models have a large number of parameters and slow detection speed, which cannot meet the widespread application requirements of the autonomous driving industry. Therefore, a lightweight semantic segmentation algorithm based on the improvement of the DeepLabv3+ network is proposed. More specifically, the algorithm replaces the original model’s backbone network used for feature extraction with MobileNetv2 to reduce computational complexity. Furthermore, it introduces improved attention refinement and feature fusion modules in the encoding and decoding stages respectively to enhance segmentation accuracy and boundary feature extraction capabilities. The experimental results of the proposed algorithm on the Cityscapes validation set demonstrate good segmentation accuracy and faster inference speed with reduced parameters compared to the DeepLabv3+ model.