With the advancement of deep learning in the automotive domain,more and more researchers are focusing on autonomous driving.Among these tasks,free space detection is particularly crucial.Currently,many model-based app...With the advancement of deep learning in the automotive domain,more and more researchers are focusing on autonomous driving.Among these tasks,free space detection is particularly crucial.Currently,many model-based approaches have achieved autonomous driving on well-structured urban roads,but these efforts primarily focus on urban road environments.In contrast,there are fewer deep learningmethods specifically designed for off-road traversable area detection,and their effectiveness is not yet satisfactory.This is because detecting traversable areas in complex outdoor environments poses significant challenges,and current methods often rely on single-image inputs,which do not align with contemporary multimodal approaches.Therefore,in this study,we propose a CFH-Net model for off-road traversable area detection.This model employs a Transformer architecture to enhance its capability of capturing global information.For multimodal feature extraction and fusion,we integrate the CM-FRM module for feature extraction and introduce the novel FFX module for feature fusion,thereby improving the perception capability of autonomous vehicles on unstructured roads.To address upsampling,we propose a new convolution precorrection method to reduce model parameters and computational complexity while enhancing the model’s ability to capture complex features.Finally,we conducted experiments on the ORFD off-road dataset and achieved outstanding results.展开更多
文摘With the advancement of deep learning in the automotive domain,more and more researchers are focusing on autonomous driving.Among these tasks,free space detection is particularly crucial.Currently,many model-based approaches have achieved autonomous driving on well-structured urban roads,but these efforts primarily focus on urban road environments.In contrast,there are fewer deep learningmethods specifically designed for off-road traversable area detection,and their effectiveness is not yet satisfactory.This is because detecting traversable areas in complex outdoor environments poses significant challenges,and current methods often rely on single-image inputs,which do not align with contemporary multimodal approaches.Therefore,in this study,we propose a CFH-Net model for off-road traversable area detection.This model employs a Transformer architecture to enhance its capability of capturing global information.For multimodal feature extraction and fusion,we integrate the CM-FRM module for feature extraction and introduce the novel FFX module for feature fusion,thereby improving the perception capability of autonomous vehicles on unstructured roads.To address upsampling,we propose a new convolution precorrection method to reduce model parameters and computational complexity while enhancing the model’s ability to capture complex features.Finally,we conducted experiments on the ORFD off-road dataset and achieved outstanding results.
