Detection Of Traversable Areas Of Roads Based On Hyperspectral Images

Environment perception is the core technology of the autonomous driving system,and the passable area detection is an important part of environment perception,which is of great significance to the path planning of autonomous driving.Hyperspectral images have been widely used in the field of remote sensing and target detection because of their wide range of spectral sensing and including physical properties related to the material of the sensing object.This paper mainly studies the passable area detection based on hyperspectral images,uses the material information of the perceptual objects contained in the hyperspectral image to solve the problems in the passable area detection based on RGB images,and realizes the passable area by segmenting the hyperspectral image.The main contents are as follows:(1)The road passable area detection based on RGB image is studied.In this paper,3799 road images are collected using the color camera on the unmanned driving platform,and the passable area is marked.The image contains urban and rural scenes,accounting for40% and 60%,respectively.The road surface material includes asphalt,cement,sand,gravel and soil,ensuring the diversity of the data.At the same time,for different scenarios,a classic semantic segmentation network was trained on the collected images,and the segmentation effects of different models were tested.In addition,for the segmented road,a road direction detection method based on the vanishing point is proposed and used for curve analysis.(2)A hyperspectral data set for road segmentation tasks is constructed.In order to use hyperspectral images for road detection,this paper constructs a hyperspectral road image dataset based on the previous work,collects data with a color camera and hyperspectral cameras in the same scene,and annotates them.The data set size reaches 3799 groups.In addition to three-channel RGB images,each scene also contains 25-channel near-infrared band images and 16-channel visible light band images,ensuring the spectral diversity of the data.In addition,classic semantic segmentation networks are trained on the constructed data set,and the segmentation effects of different models are tested,providing a standard data set for the hyperspectral image road segmentation work.(3)A semantic segmentation model based on band selection is proposed.The hyperspectral image contains dozens of bands,and there may be feature redundancy or unwanted noise.In order to solve this problem,this paper introduces a band selection module based on sorting.Firstly,the number of bands with the best effect is verified.Then,we compare proposed model with other road segmentation models based on hyperspectral images.The experimental results show that the band selection module improves the segmentation performance.Finally,we compare our model with the model that uses the RGB image of the same scene as the input.The proposed model performs better in unstructured scenes,and the experimental results verify the effectiveness of the band selection module and the application value of hyperspectral images in road segmentation tasks.