Research On Unstructured Road Drivable Area Recognition Method Based On Deep Learning

With the rapid development of computer vision technology,vision-based environment perception systems have gradually become a research hotspot.The recognition of drivable areas is one of the key technologies in environment perception systems,especially the recognition of drivable areas on non-structured roads,which has significant implications.Only by quickly and accurately identifying the drivable areas on non-structured roads can autonomous driving be truly realized.This paper conducts indepth research on the recognition method of drivable areas on non-structured roads based on deep learning technology,and the main research contents include:First,a detailed analysis was conducted on the recognition of non-structured roads and the identification of drivable areas on these roads based on vision and deep learning.The characteristics of non-structured roads and the existing problems with various recognition methods were summarized,such as the absence of lane markings and unclear road boundaries.To address these issues,a deep learning-based method was proposed for the identification of drivable areas on non-structured roads,which uses the Bilateral Segmentation Network as the base network model for identifying drivable areas on nonstructured roads.Two improvements were made to Bi Se Net.The main network was replaced with a short-term dense connection network to reduce the model’s computational complexity and improve its recognition speed.A unified attention fusion module was added,whose channel and spatial attention can enrich the fused feature representation and improve segmentation accuracy.The model was trained on a constructed non-structured road dataset in a campus environment.The model’s feasibility and effectiveness were verified using common evaluation metrics.The proposed method for non-structured road drivable area recognition was validated by selecting a typical non-structured road scenario,the underground mining slope road.A dataset was constructed using image data of the underground mining slope road captured by a front camera of a mining truck,and data augmentation was applied to expand the dataset.The drivable area of the underground mining slope road was labeled using the labelme image annotation tool.The labeled image data was fed into the network and trained with environmental settings and hyperparameters.The performance of the model was evaluated using common evaluation metrics and compared with the results of three other classic networks.The results showed that the proposed method is feasible and superior.Finally,on the basis of identifying the drivable area of unstructured roads,the midline fitting of the drivable area is studied.In order to extract the contour of the drivable area of the unstructured road,the obtained drivable area is preprocessed.According to the image center moment theory,the center point coordinates of the drivable area are calculated by using the image center moment algorithm.Finally,the least square method is used to fit the middle line of the road.The results show that the middle line of the road is obtained by fitting the middle line of the driving area.The research in this paper combines deep learning with computer vision,effectively solving the problem of identifying drivable areas on unstructured roads and providing useful references for environment perception in autonomous driving vehicles.