Research On Wild Road Extraction Technology Based On Semantic Segmentation

To improve vehicle safety and driving experience,artificial intelligence-based automatic driving technology has begun to assist and replace manual driving.As the core of autonomous driving technology,although road detection technology performs well on urban roads dominated by structured roads,its robustness in wild still needs to be improved.As the extremely special case of unstructured roads,wild roads have complex surface features,blurred boundaries,and background features similar to road surfaces.The mainstream semantic segmentation network suitable for urban roads often performs poorly on wild roads.In this thesis,road recognition accuracy is improved by enhancing the ability of the model to capture detailed features and context information in combination with the characteristics of the wild road.The specific research contents are as follows:(1)In view of the fact that there are few publicly wild road segmentation datasets and rarely consider the influence of environmental interference such as seasons and light,this thesis constructs a wild road semantic segmentation dataset which consider the above factors and makes it public.The dataset contains 8011 wild road pictures in different seasons and light intensities,and manual labeling of all images has been completed.(2)In view of the poor performance of mainstream semantic segmentation methods in wild road,this thesis proposes the MICNet.In MICNet,the Information Sharing Layer facilitates the shallow information exchange of the feature parameters of the dual-stream network;The Multi-information Concatenate Module alleviates the problem of shallow information being lost in the deep network;The Two-Path Way Semantic Inference Layer is used to enhances the context information extraction ability of the network;The Detail guidance module facilitate the network’s capture of road detail features.The MICNet is compared with some state of the art on the local server.Its advantages in the wild road segmentation task are verified with 89.5% Io U.(3)To verify the practicability of the wild road dataset and the MICNet,this thesis realizes the conversion of the input image and the deflection angle required for the car to travel along the road on the smart car.First,the distortion correction algorithm is used to solve the image distortion problem caused by the camera quality.Then,the undistorted image is sent to the Tensor RT accelerated model to quickly predict the wild road result.Finally,the perspective transformation method is used to map the target points on the road centerline of the segmentation result from the head-up perspective to the top-down perspective to improve the calculation accuracy of the deflection angle.In the real wild environment test,it takes about 0.04 seconds to calculate the transformation of each picture frame into a deflection angle,which meets the real-time requirements.