Cascaded residual attention mechanism for semantic fusion in vehicle driving
Xiaohang Li, Jianjiang Zhou · 2023
Multiple sensors are often used to work together in vehicle driving, but how to effectively fuse the data of each sensor is a difficult research point. The attention mechanism can assign different weights to each target, allowing more computing resources to focus on key targets, greatly improving computational efficiency and accuracy. In this paper, ResNet34 and SalsaNext are used as encoders of dual-stream networks to extract general features of images and point clouds respectively, and a cascade residual attention fusion strategy is proposed, which is used between two-stream networks to fuse features from two modalities at different encoding stages. Experiments on the SemanticKitti dataset show that this fusion strategy has better performance than single-stage fusion and PMF fusion structures.