Research On Environmental Perception And Risk Prediction For Intelligent Vehicle In Complex Environment

Vehicle intelligence is a crucial direction for the development of the global automotive industry,and the advancement of intelligent vehicles presents a significant historical opportunity for our nation to transform from a major automotive country to a powerful one.Currently,high-level autonomous driving technology has yet to satisfy application demands.Issues including environment perception and driving risk prediction are the biggest hurdles limiting safety.Accurately perceiving the environment and predicting potential risks can directly ensure the driving safety of intelligent vehicles,improving the performance of these two areas also presents significant bottlenecks to the development of high-level autonomous driving.At present,intelligent vehicles can achieve high environmental perception accuracy in relatively simple traffic scenarios,but with increasing road complexity,the performance of their environmental perception algorithms significantly decreases.Therefore,accurate environmental perception and driving risk prediction in complex road traffic scenarios are key issues addressed in this article.The primary research approach focuses on the philosophy of “front-end environmental perception-back-end risk prediction”,systematically studying methods of environment perception in intelligent vehicles from various perspectives including high-precision localization,object recognition and tracking,and accordingly implementing driving risk predictions.The main reason for the lack of environmental perception accuracy is the limited performance of a single sensor,which reduces the adaptability of intelligent vehicles to complex traffic scenarios.These limitations can be effectively compensated by utilizing multi-source data fusion theory,which enhances the sensing effects of intelligent vehicles and improves the stability against external interference.In this study,we conduct in-depth research on object level fusion and feature level fusion,proposing a high-precision localization method of object level fusion based on road scenario models and a multi-object detection and tracking method based on feature level multi-data fusion.We propose a driving risk prediction method based on vehicle trajectory prediction,aiming to provide theoretical support for the safe driving control of intelligent vehicles.The specific research content is as follows:(ⅰ)An intelligent driving experimental platform has been built for real vehicles.The theory of multi-data fusion is studied,and the fusion method suitable for complex traffic scenario environment perception is analyzed.To address the issue of most calibration methods deriving from engineering measurements,hence reducing calibration precision,we proposed an external parameter calibration method for cameras and lidar based on a chessboard plane geometric alignment,replacing point alignment with surface alignment,and converting the engineering problem of multi-sensor calibration into a planar geometric alignment and direction vector solving problem.This method simplifies calibration steps,improves the utilization rate of the lidar point cloud data,and enhances the precision and stability of joint calibration.This system provides platform support for environmental perception and driving risk prediction in complex scenarios.(ⅱ)A high precision localization method is proposed based on a multi-level road scene model of the "feature-3D-location" components.The multi-level road scenario representation model is designed to uniquely represent road scenes with the minimal data storage space by means of multi-dimensional features.We propose a double-scale topological model which fuses self-motion estimation and previous localization results.This model is replaced to GNSS coarse localization.This model also achieves the purpose of narrowing the localization range and improving the localization efficiency.In response to the instability of lidar point cloud feature extraction,we introduced an approach for image global feature extraction where lidar point cloud is rasterized into an image-style lidar point cloud.We proposed a laser-image global feature extraction method to achieve laser lidar feature matching and image feature matching,obtaining localization results under two different methods.An object level fusion localization method is then proposed to merge lidar localization candidate points with image localization candidate points,enhancing localization precision.Localization precision can reach 9 cm — an improvement of 7 cm compared to other methods —providing an accurate localization reference for multi-object detection and driving risk prediction in complex scenes.The proposed method solves the problem of low localization precision in complex traffic scenarios such as blind area.(ⅲ)A multi-object detection and tracking method is proposed based on feature-level fusion of multi-data.The feature level fusion method of laser point cloud and image is proposed to solve the insufficient generalization ability of single sensor.We introduced a new network-RobustFusionNET,targeting illumination changes,rainy or foggy weather,smog,or nighttime,aiming for improved multi-object detection performance in complex traffic conditions.In this network,we propose a point-by-point alignment fusion technique based on K-means++ clustering.It first clusters image features accurately before fusing them with point cloud features,yielding corresponding image features.This ensures precise fusion results even in extreme conditions when point clouds are particularly sparse.An adaptive correction quantity is added to the pixel coordinates of the projective point cloud to handle errors brought by joint calibration.We also employ a multi-object tracking method based on third-order Kalman filtering using millimeter-wave radar,which can discard false objects,thereby enhancing tracking precision.This approach improved the detection accuracy by 1.01%compared to other methods.The proposed method solves the problem of multi-object detection in complex traffic scenarios such as harsh weather.(ⅳ)A driving risk prediction method is proposed based on vehicle trajectory prediction.On the basis of the output vehicle prediction trajectory,a driving risk domain considering driver behavior uncertainty is used as a measure of risk probability.Taking into account environmental events that have a significant impact on vehicle safety as a measure of severity,we achieve a quantified perceived risk that adapts to the uncertainty of complex driving scenarios.Based on real vehicle trajectory data and the theory of Bayesian probability,we weight and fuse the quantified perceived risk at all times in the prediction interval,ultimately achieving a prediction of potential collision risks in future driving.The study predicts collision risk based on the vehicle position information output by the predicted trajectory,and the experimental results in multiple scenarios show that the risk prediction moment of this method can be about 0.9s earlier compared to the current authoritative risk prediction indicators.The proposed method overcomes the problem of poor real-time risk prediction in comprehensive vertical and horizontal complex scenarios.The aim of this method is to enhance the safety of autonomous vehicles and provide theoretical support for the commercialization of high-level autonomous driving.This study focuses on addressing the challenges faced by intelligent vehicles in complex traffic scenarios,such as low localization precision,difficulty in detecting object in harsh weather,and poor real-time prediction of comprehensive longitudinal and lateral risks.Through theoretical and methodological innovations,on-road experiments,open datasets,and simulated experiments,significant progress has been made.The research findings provide a theoretical and technological foundation for our country to grasp the global leading core technologies and influence in the field of intelligent automotive industry,which is crucial for promoting the application of intelligent vehicle safety technology in complex traffic scenarios.