Research On Path Planning Method For Unmanned Ground Vehicle Guided By Overhead Image

The path planning technology of Unmanned Ground Vehicles(UGV)is one of the core issues in the field of autonomous driving technology research.The existing path planning methods have strong dependence on environmental prior information and precise localization,and lack of guidance in the path searching process.In order to ensure path planning safely and reliably for unmanned ground vehicles in complex environments with weak prior information,such as battlefield operation environment and desolate search support environment,an innovative path planning technique based on overhead image guidance,environmental inituition information and accurate model driving is proposed.The main contents and innovations of this paper are as follows:1.An automatic local guidance direction reasoning method based on satellite map/aerial image is designed.Firstly,the semantic segmentation network is used to extract roads,then the road skeleton is obtained by morphological post-processing methods and the local guidance direction of the vehicle is deduced according to the real-time position of the vehicle,the local road skeleton model and the global target guidance point.This method can not only reduce the cost consumption and avoid the great subjectivity brought by manual route drawing,but also improve the autonomous decision-making ability.2.A driving intention prediction method is explored.The method takes the local guidance direction and local visual perception information as input,learning human driving experience from the historical data by using Generative Adversarial Nets.Then an intention region would be predicted to provide reference for subsequent local path planning.This method efectively overcomes the problem of intermediate fusion between perceptual data information and logical reasoning in end-to-end navigation.Experimental results show that this method has a perfect performance of generating driving intention region,which can provide reference for the subsequent path planning reasoning process.3.A local path planning method combined with the prediction results of intention region is implemented.Firstly,the local environment model is obtained by combining the prediction results of intention region with the accurate environmental model obtained by other sensors;Then,a method based on cubic spline curve is used to generate candidate path set;In the stage of optimal path selection,a function is designed which fully considers vehicle collision avoidance and intention area prediction results to make the search more efective while ensuring safety and improve the real-time performance of the algorithm.Experimental results show that this method can quickly and efectively generate safe and smooth local path.