ABSTRACT
A convolutional neural network segmentation method is proposed to meet the real-time requirements of automated driving systems and the problem of fuzzy target boundary segmentation and inaccurate segmentation of similar objects caused by pooling and convolution operations in current semantic segmentation methods of traffic scenes. A model with an attention mechanism has been proposed to detect obstacles in traffic scenes. The model constructs a feature extraction block and introduces various extended convolutions in the feature extraction process to extract more detailed image feature information. After the information is decoded, dual-channel and spatial attention modules are added, and functional information of different scales is combined during up-sampling, making the semantic information richer. The test results show that the proposed lightweight semantic segmentation model has a mean intersection ratio (MIoU) of 71.5% on the Cityscapes dataset, achieving good obstacle detection results, and meeting the real-time requirements of autonomous vehicles.
