Road Obstacle Perception And Parametric Analysis Based On Image And Point Cloud Fusion

At present,the period of artificial intelligence has come,and the manufacturing level of various sensors has made considerable progress.The level of intelligent vehicles has been continuously improved,and autonomous driving has gradually changed from fantasy to reality.As a key part of vehicles to obtain road information,the performance of perception system is directly related to the safety of autonomous vehicles.At present,the demand for perception of autonomous driving is evolving from 2D space to 3D space,which requires the perception system not only to get the location of key road targets,but also to know the pose of the target.Current problems are reflected in two aspects,one is how to realize accurate estimation of target parameters,the other is how to ensure the real-time perception.Based on these two problems,this article studies from the direction of image,point cloud and fusion,and proposes a fusion-based perception algorithm,which realizes real-time accurate perception and parameterized analysis of road obstacles.The main research contents and achievements of this paper are as follows: Firstly,this paper studies the image-based visual perception algorithm,designs a Backbone suitable for object detection and semantic segmentation,and builds a target detection and semantic segmentation model on this basis.Through feature sharing,this paper builds a joint model,which greatly reduces the amount of parameters and calculations.Experiments show that the joint model can perform target detection and semantic segmentation tasks simultaneously at a faster speed.Under the premise of ensuring accuracy,the requirements of real-time are met;Then,based on Point Net,this paper designs a network named Point SSD that directly processes the raw point cloud data,which can be applied to point cloud target classification,semantic segmentation,and 3D object detection tasks.Because of the attention mechanism introduced in the architecture design,Point SSD can extract more effective features with a lighter structure;Finally,this paper proposes a Frustum-Point SSD object detection algorithm based on the fusion of image and point cloud.Based on the joint network and Point SSD,the algorithm realizes accurate parameter estimation of road obstacles.By merging the semantic features of the image point-by-point,the problem of point cloud sparseness is compensated,and the accuracy of small and difficult targets is improved.In the process,the point cloud data organization format is optimized to increase the speed of inference.On the KITTI 3D object detection data set,the algorithm shows a high accuracy,with NVIDIA GTX1080 Ti as the computing platform,the fusion algorithm inference time is 54 ms.