Semantic segmentation of drivable region for unstructured roads
Accurate identification of the unstructured road driving regions can provide theoretical basis for local autonomous decision-making and route planning of unmanned combat for military intelligent vehicles.In view of the fuzzy edges and similar features of the unstructured roads,existing segmentation algorithms are difficult to meet the requirements of the military intelligent vehicles'accuracy recognition.In this paper,attention Transformer DeepLabv3+(ATD),a semantic segmentation algorithm for driving area oriented to unstructured roads,is proposed.Convolutional Block Attention Module(CBAM)module cascaded in coded feature extraction,which enhances the adaptive pixel weights of semantic information in different channels and spatial dimensions and strengthens the feature coding ability in complex environments without losing the accuracy of the semantic segmentation and recognition.The introduction of Transformer Multi-Head Attention in decoding to strengthen the relevance of spatial location information and realize the fine-grained reasoning of unstructured road edges.Based on about 6 000 unstructured road data sets built by ourselves,the experimental results show that MPA and MIoU are improved by 5.65％and 4.20％on average compared with the network model.
NETL
NSTL
万方数据
