Road Detection In Off-Road Environment

In recent years,self-driving technology has been widely used in more and more scenarios.In off-road environment,self-driving technology is the basis of military unmanned ground vehicles(UGVs).As an important part of unmanned systems,road detection is a necessary support for autonomous navigation,which plays a vital role in environment perception and path planning.Unlike lane detection in urban environment,roads in off-road environment are more complicated.This paper makes a systematic study on road detection in unstructured environment and its application in local path planning.The main contributions and innovations of this paper include the following two points:Firstly,in terms of engineering application research,we apply road detection technology in autonomous navigation of UGV in unstructured environment.We apply an existing road model based on discrete roadside points representation and use an end-to-end convolutional neural network(CNN)to make road prediction.Then the algorithm is integrated into the self-driving system and uses the RGB picture collected by the vehicle camera in real time as input.Besides,we propose a local path planning method based on visual road boundary detection.We firstly accomplish the camera internal parameters calibration and the external parameters calibration between the camera coordinate system and the planning coordinate system.Homography matrix between the image plane and the planning plane is obtained after calibration.By projecting the roadside points detected in the image to the planning coordinate system,the UGV can perform robust local path planning without precise position information.According to the actual situation,optimization strategies are proposed,such as correcting homography matrix by combining real-time vehicle body postureinformation and optimizing projection results by multi-frame fusion.Secondly,in terms of theoretical research,in order to overcome the problem that the existing road models based on discrete roadside points are not enough to represent complex roadsides,we propose a roadside representation model based on trigonometric function and a local roadside coordinate system.First,based on the labeled roadside points,we constructed a local roadside coordinate system for both the left and right roadsides.In the local coordinate system,trigonometric function is used to perform curve fitting on the labeled roadside points to obtain the ground truth trigonometric function representation of the roadside.Then,a convolutional neural network is designed to perform end-to-end prediction.By using the trigonometric function as the ground truth and designing the corresponding loss function,the network can predict trigonometric function coefficients.Our model can directly obtain continuous roadside function expressions,and handle more complicated roadside situations.A series of comparative experiments are conducted on our self-built road datasets in off-road environment,and the experimental results show that the road model detection method based on trigonometric functions proposed in this paper performs well in road detection in the off-road environment.