Vehicle Pose Detection And Its Semantic Map Application

Vehicle pose detection is an important research content in the field of smart driving perception.Vehicle pose information is one of the important basis for driving decisions.A semantic map refers to a map containing the semantic information required by the smart driving system.The semantic information refers to information that can be directly used by the smart driving system,and is different from information used for human use in a traditional map.This paper proposes a new convolution method—grille convolution,to achieve accurate and rapid vehicle attitude detection,and applies the vehicle pose as dynamic semantic information in the semantic map system.Aiming at the problem of vehicle attitude detection,this paper uses the deep learning method,improves the traditional attitude detection network by using partial connection network,solves the difficulty caused by the perspective effect on vehicle orientation detection,and improves the effect of the attitude detection.In order to solve the problem of lack of training data for vehicle attitude detection,this paper builds a set of deep learning training data generation system for automatically generating training data with vehicle attitude tags in order to realize network training.Based on the partial connection structure,a partial connection attitude detection network(LC-Pose)is proposed and training,parameter tuning and testing are performed on the generated data set.Compared to the traditional attitude detection network,the average angle similarity(AOS)index is obtained.More than 10% increase.This paper continues to analyze the advantages and disadvantages of some connected attitude detection networks(LC-Pose),and proposes a cooperative network to reduce the performance sacrifice between the detection part and the pose estimation part.In order to improve the performance of partially connected networks,this paper innovatively proposes grille convolution,which can satisfy the processing capacity of perspective ambiguities and reduce the possibility of overfitting the model.Grille Net is designed based on grille convolution to further improve the vehicle attitude detection effect.Tested on the world’s largest smart driving data set KITTI and compared with the top ten methods in the world,the average precision(AP)of 86.46% and the average orientation similarity(AOS)of 88.59% were obtained.In terms of system,this paper designs a lightweight,easy-to-expand smart driving semantic map software framework that can meet the needs of various smart driving applications.This paper analyzes the semantic information representation in semantic maps and semantic information storage methods,and uses hierarchical map expressions to express and store the semantic information in semantic maps,and designs a unique structure suitable for smart driving situations.Finally,combined with the previous GPS gesture detection network of the Grille Net and the driving data set of Changchun,a semantic map with vehicle posture semantic information was created to implement the semantic map application.The vehicle-pose-detection-network studied in this paper provides a powerful information basis for smart driving decisions and target vehicle behavior prediction.The semantic map system provides a comprehensive and efficient information acquisition platform for smart driving.Combining dynamic semantic information such as vehicle posture semantic information can achieve efficient smart driving decisions and macro global traffic behavior planning and forecasting.