Research On The Driving Risk Assessment For Intelligent Off-Road Vehicle Based On Terrain Parameter Fusion Estimation

As a technology that can improve driving safety and reduce traffic accident rate,intelligent vehicle technology has now become the strategic development direction of the world’s automobile industry.China’s strong promotion of intelligent vehicle industry also provides an opportunity for the application and development of intelligent off-road vehicle technology in agriculture,forestry,military,mining,and other fields.It is the basic requirement of the intelligent off-road vehicle to achieve the task of safe driving on complex and changeable terrain conditions by obtaining more abundant environmental information.Based on this,this paper focuses on the driving risk assessment of intelligent off-road vehicles based on terrain parameters fusion estimation.This research is relying on the university-enterprise cooperation project "Intelligent Terrain Integrated Management and Intelligent Chassis State Information Estimation Technology Development",and the Jilin Provincial Natural Science Foundation project "Research on Key Technologies of Intelligent Tire and Information Fusion for Off-road ground Feature Perception".This research is about terrain parameters fusion estimation in complex cross-country condition,and vehicle state prediction and risk assessment with uncertainty.Multi-modality data includes colored photos,3D point clouds from lidar,and vehicle dynamic states.Terrain mechanical parameters and geometric parameters are estimated with parameter estimation method and data fusion theory.Driving risk is finally assessed with parameter uncertainty,with provides trafficability prediction result for intelligent off-road vehicle.In this paper,multi-source data collection and feature extraction of typical off-road scenes are firstly carried out.Existing public data sets are not complete,failing to support the research of this paper.So,we set up a data collecting vehicle platform,and design the mechanism for multi-source data synchronization sampling.We conduct data collection work and build a multi-source databases containing various cross-country scenes including sand,dirt,asphalt,and snow terrain.Based on the multi-source database,the requirements of terrain classification are further analyzed,and the dynamic features and visual features contained in the multi-source data are calculated to provide data support for the subsequent research on terrain classification algorithms.Secondly,the cross-country road fusion classification method based on vehicle dynamics and vision is studied.In order to realize the classification of the terrain surface in front of and under the vehicle,a vision-based terrain classification algorithm is designed.Firstly.The semantic segmentation network combined with dynamic region of interest visual field model is used to realize the terrain classification considering the confounding of the road surface.Then the terrain classification algorithm based on vehicle dynamic response is studied.The terrain classification model based on LSTM is sampled,and the optimal configuration is obtained by comparing different model structures and parameters.Using the real vehicle data as the validation data set,the results show that the terrain classification algorithm proposed has high classification accuracy.According to the data fusion theory,the fusion requirements are analyzed in detail,and a fusion algorithm for under-vehicle terrain classification is built.In order to improve the reliability of the fusion algorithm,a hidden Markov class tracking algorithm is proposed in this paper,so that the fusion system can produce continuous and accurate output even when classification algorithms fail.From the test results,the proposed under-vehicle terrain fusion classification algorithm with class tracking achieves the improvement of classification stability and accuracy.Thirdly,the fusion estimation of terrain mechanical parameters and geometric parameters is studied.In order to describe the terrain,this paper describes and quantifies the road surface from two perspectives: mechanical and geometric parameters.In terms of the mechanical parameters of the road surface,the adhesion coefficient estimation based on the terrain type and the adhesion coefficient estimation based on the vehicle dynamics are combined to realize the adhesion coefficient estimation.Aiming at the demand of road adhesion coefficient fusion estimation,a feature level fusion estimation method of road adhesion coefficient based on dynamic sampling semi-supervised learning was proposed.the dynamic adhesion distribution model is constructed with the adhesion estimation based on the road type and the adhesion coefficient estimation based on the vehicle dynamics.The model is verified by using real vehicle data.The road adhesion coefficient fusion estimation algorithm based on dynamic sampling semi-supervised learning can estimate with the different fusion modes correctly and provide the continuous estimation results of adhesion coefficient distribution.In the aspect of terrain geometric parameter estimation,the slope angle estimation based on vehicle state and the slope angle estimation based on laser point cloud are carried out.The data-level fusion mechanism between them is established to finally get the slope angle distribution and obstacle information of the terrain ahead.Through the simulation platform and the real vehicle data,the proposed algorithm is verified.The results show that the algorithm can estimate the terrain slope angle information correctly.Finally,the longitudinal risk assessment of intelligent off-road vehicles considering the uncertainty of terrain parameters is studied.On the premise of considering the uncertainty of terrain parameters,the risk assessment of intelligent off-road vehicle longitudinal driving is carried out with the uncertainty framework.Based on the theory of vehicle trafficability,the driving risk is preliminarily analyzed,and the corresponding failure indexes are established based on three failure modes: stability failure,traction/braking force failure and clearance failure.According to each failure form,a driving risk assessment framework considering the uncertainty of road parameters is established,the uncertainty sources are quantified,the propagation route of uncertainty is analyzed,and the probabilistic reasoning method of failure index is established for each failure index successively.According to whether uncertainty propagates over time,static driving risk assessment and dynamic driving risk assessment are used to carry out risk assessment,and the calculation process of different driving risks is integrated and unified scheduling,to realize the integrated vehicle longitudinal driving risk assessment.In order to verify the rationality and effectiveness of the algorithm,simulation test and offline test based on real vehicle data are carried out respectively.In the simulation test,the co-simulation test environment is built,and the simulation test is carried out in combination with the automated test method.The key working conditions are analyzed to enhance the working conditions where the risk assessment algorithm is prone to failure and increase the test credibility.In the off-line testing process based on real vehicle data,an off-line testing platform capable of synchronizing multi-source data was built,and the off-line verification was carried out using multi-source data in real off-road scenarios,and a rolling time-domain testing method was designed to carry out automatic off-line testing of the algorithm.Simulation and off-line testing verify the accuracy of the proposed driving risk assessment algorithm.