| Society of Automotive Engineers(SAE)has clearly defined the automated driving levels into six categories from L0 to L5.Thereinto,the L4 automated driving has deeply attracted research interest due to its zero intervention from human drivers in specific scenarios.However,there are still numerous driving scenarios that occur at a small probability but in a high-risk manner,named“corner cases” [1],[2],which pose a safety threat to passengers or vehicles.These safety-oriented corner cases have covered the four systems in autonomous driving applications: perception,decision-making,planning,and control.In this dissertation,the four-wheel-independently-actuated(FWIA)chassis configuration is introduced,and it is integrated within the autonomous vehicle.Although all the corner cases cannot be discussed in this dissertation,several corner cases are analyzed and can be extended to other similar safety-threat scenarios.Five challenging problems related to these corner cases can be listed as follows:(1)Perception—Multi-Li DAR Fusion: A systematic solution of multi-Li DAR data processing is one cornerstone of autonomous driving,including calibration,segmentation,filtering,clustering,and identification.Due to the costly Li DAR sensors,most commercial solutions adopt single Li DAR for object detection.However,the detection performance too limited to possibly cause the misiden-tification for on-road objects since the single-source point-cloud data is biased and less robust.A multi-Li DAR solution can deal with these problem in commercialization,but its challenges in imple-mentation are fourfold: i)How to guarantee a better accuracy and robustness than a single-Li DAR solution.ii)How to filter noise and outliers with an advanced method.iii)How to improve the clus-tering performance based on the principle of laser reflection.iv)How to accurately extract objects with the same data format yet from various Li DAR data sources.(2)Perception—Radar-Vision Fusion under Li DAR Failure and Various Illumination: When the Li DARs suffer from power outage or functional failure,millimeter-wave radars and cameras should perform the task of environmental perception in automated driving.Even worse,cameras are incredibly sensitive to lighting conditions that means their detection performance will be poor under lower illumination.For example,self-driving at nights.This common yet critical scenario can be extended to a more general case through darkened tunnels that suffer from frequent illumination changes.How to develop a radar-vision fusion solution becomes the key to widely apply self-driving vehicles,particularly for robotaxies.(3)Decision-Making—Learning on Human Driver Behavior: A driving scenario is prede-fined as the ego autonomous vehicle suffers from an oncoming vehicle with high beams on an un-structed road.In order to imitate the driver-like behavior,the stochastic end-to-end model should be trained based on the collected driving data.Human-like driving behavior learning has the capability of avoiding the human-machine trust problem.Personalized driving strategies should be also learned and identified by the human-like learning.It provides an example for learning human safe driving actions in the scenario that can be generalized to other safety-related scenarios.(4)Planning—Low-Cost Velocity Estimation and Collision-Free Path Replanning: Vehicle lateral velocity is frequently measured by differential global positioning systems(DGPS)that are costly in real implementation.A low-cost velocity estimation method is of practical importance to plan an evasive path,and further control the vehicle motions.Besides,path replanning in extreme driving conditions should not only aim at generating available collision-free path candidates,but also consider the enhancement of safe road capacity,where the conventional single mass-point model has limitations due to its neglect of vehicle dimensions.(5)Control—Safety-First-based Post-Impact Control: According to traffic accident statistics,a higher death rate imputes secondary crashes than first impacts.Numerous researchers have made efforts to improve the vehicle stability to avoid the first collision.However,the solution of secondary collision mitigation should be based on the safety-first motion control rather than the conventional concepts of vehicle stabilization.How to transform the first-impact energy and how to fully utilize the generalized tire forces become the core concepts of compensatory predictive control with the FWIA configuration.To tackle these challenging problems,the research work for the typical safety-oriented scenarios in L4 automated driving can be separated into fivefold accordingly.(1)Systematic Multi-Li DAR Fusion Solution: In contrast to the single Li DAR system,multi-Li DAR sensors may improve the environmental perception for automated vehicles.However,an elaborate guideline of multi-Li DAR data processing is absent in the existing literature.A systematic solution for multi-Li DAR data processing that orderly includes calibration,filtering,clustering,and classification is designed.As the accuracy of obstacle detection is fundamentally determined by noise filtering and object clustering,a novel filtering algorithm and an improved clustering method are proposed within the multi-Li DAR framework,respectively.To be specific,the filtering approach is applied based on the occupancy rates(ORs)of sampling points.Besides,ORs are derived from the sparse “feature seeds” in each searching space.For clustering,the density-based spatial clustering of applications with noise(DBSCAN)is improved with an adaptive searching(AS)algorithm for a higher detection accuracy.More robust and accurate obstacle detection can be further achieved by combining AS-DBSCAN with the proposed OR-based filtering strategy.An indoor perception test and an on-road test were conducted on a fully instrumented autonomous hybrid electric vehicle.Experimental results have verified the effectiveness of the proposed algorithms,which facilitate a reliable and applicable solution for obstacle detection.(2)Radar-Vision Fusion under Li DAR Failure and Various Illumination: Autonomous ve-hicles are widely equipped with radars,camera,and Li DARs due to their complementary capabilities of environment perception.However,it becomes a critical task to accurately track the trajectory of the preceding vehicle when the host vehicle suffers a Li DAR failure under challenging lighting condi-tions.An integrated learning-based solution is proposed to address this critical issue.It is composed of a Q-learning-based Gaussian mixture model(QLGMM)for clustering dense radar data,a weight-scheduled method for radar data association,and a switchable recurrent neural network(RNN)with dual-level long short-term memory(LSTM)cells for trajectory tracking under Li DAR failure and var-ious illumination levels.Specifically,the QLGMM is first introduced to improve the conventional GMM-EM algorithm with the cluster number determined by a Q-learning approach.Subsequently,the weight-scheduled method is presented to associate the data from multiple radars.Furthermore,a switchable dual-level LSTM(SDL)network is developed to adaptively fuse the trajectories from the radar and camera streams based on three lighting modes(namely day mode,dusk mode,and night mode).The training data and testing data were acquired on a fully-instrumented autonomous vehicle.Experimental verification demonstrates that the proposed method can achieve a promising improvement for Li DAR fault-tolerant trajectory tracking under different lighting conditions.(3)Decision-Making based on Driver Behavior learning: Oncoming vehicle high-beams im-pair the detection performance of cameras in autonomous driving applications.In this scenario,mod-eling the stochastic human driving behavior becomes an essential and challenging task for vehicle safety enhancement.Based on learning from human driver data,a novel and integrated methodology that generates driving operations of suffering oncoming high-beams is investigated.By decompos-ing the human drivers’ pedal and steering responses,a parallel autoregressive input-output hidden Markov model(p-AIOHMM)is introduced to capture the temporal dependencies of the decomposed driving actions.The p-AIOHMM that consists of two feedback AIOHMMs has the capability of imitating the stochastic actions of human drivers.Moreover,an improved Wasserstein generative adversarial network method with gradient penalty(WGAN-GP)is proposed to further reconstruct the pedal positions and the steering angles from the p-AIOHMM-based action probabilities.The proposed WGAN-GP algorithm includes two encoder-decoder-based generators and two decoder-based discriminators in a parallel manner.All the parameters can be learned from the real driver data collected on the autonomous vehicle platform.Experiment results have verified that the developed p-AIOHMM-WGAN-GP solution can perform a better task of driving behavior generation when suf-fering oncoming high-beams.In handling the potential safety hazards,it can also achieve a promising improvement for modeling the individualized driving strategies.(4)Low-Cost Velocity Estimation and Path Replanning for Safe Road Capacity Enhance-ment: Based on adaptive complementary filtering(ACF)principles,a cascaded estimation method is presented to estimate the longitudinal and lateral velocities of a four-wheel-independently-actuated(FWIA)electric ground vehicle(EGV).The observation process is first carried out to estimate the lon-gitudinal velocity,followed by the lateral velocity estimator in another ACF.Both of the ACFs are regulated by a high-pass filter and a low-pass filter with adaptive filtering parameters.Quick-response motor torques are attainable to induce the relevant vehicle states in a dynamic inverse coupled tire model(ICTM).Meanwhile,an additional kinematics-based approach is lumped into ACF to enhance the robustness against modeling discrepancy to zero.The errors of estimation are proved to asymp-totically converge to zero via the Lyapunov method.Two maneuvers were conducted on an FWIA EGV platform to evaluate the proposed method.Experimental results indicate that the designed esti-mators are capable of matching with the measured longitudinal and lateral velocities accurately,and highlight it as a low-cost solution in practice.In addition,a novel hierarchical model predictive control(MPC)method is investigated for FWIA autonomous vehicles(AVs)with emergency collision avoidance,where an artificial potential field(APF)-based nonlinear MPC path replanner and a feedback compensation control(FCC)-based linear-time-invariant(LTV)MPC path follower are designed.Both replanning with circle decomposi-tion of vehicle shape,and tracking with tire force maximization,are considered simultaneously to en-large the reachable zone of path replanning and following,particularly in much aggressive situations,where the trajectories are not feasible with the conventional approaches.By virtue of the proposed method,ample space and sufficient time are available to steer appropriately and accelerate/brake inde-pendently in such hazardous scenarios.Besides,a shorter predictive horizon is introduced to evaluate both methods in a more extreme situation.The simulations modeled in the Carsim-Simulink joint plat-form demonstrate that the proposed approach can further improve path-replanning reachability and path-following safety in emergency collision avoidance scenarios,even in a shortsighted prediction.In other words,the safe road capacity for autonomous driving can be significantly enhanced.(5)Safety-First-based Post-Impact Compensatory Predictive Control: The trajectory track-ing control of independently actuated autonomous vehicles after the first impact is investigated,which aims at decreasing the secondary collision probability.An integrated predictive control strategy is presented to mitigate the deteriorated state propagation and facilitate safety objective achievement in critical conditions after a collision.Three highlights can be concluded: 1)A compensatory model predictive control(MPC)strategy is proposed to incorporate a feedforward-feedback compensation control(FCC)method.From the perspective of the definite physical analysis,it is verified that ad-equate reverse steering and differential torque vectoring render more potentials and flexibility for vehicle post-impact control? 2)With compensatory inputs,the deteriorated states after a collision are far beyond the traditional stability envelope.Hence it can be further manipulated in MPC by constraint transformation,rather than introducing soft constraints and decreasing the control efforts on tracking error? 3)Considering time-varying saturation on input,input rate,and slip ratio,the proposed FCC-MPC controller is developed to improve faster deviation attenuation both in lateral and yaw motions.Finally three high-fidelity simulation cases implemented on Car Sim-Simulink conjoint platform have demonstrated that the proposed controller has the advanced capabilities of vehicle safety improvement and better control performance achievement after severe impacts. |
