Autonomous vehicles are essential for mobility in big cities， just like how elevators make high?rise buildings livable. While significant progress has been achieved over the last 15 years， there are still several remaining challenges， namely： cost， robust performance， and trust. To address these challenges， this paper discusses research at Mcity.
At present， intelligent vehicle has become the strategic development direction of the global automotive industry. The core of automotive technology and engineering is shifting from traditional hardware layer to software layer. Software defined vehicle （SDV） has become an important trend of automotive development in future. By comparing SDV with traditional vehicle， the development， the physical structure and the cyber structure as well as the technical system of SDV are proposed in this paper.
The 3D object recognition based on lidar data is a key part of autopilot system. Voxel network is a good container for extracting point cloud features， but most of the research at present on object recognition based on voxel network focuses on single?category object. In order to meet the application demand of unmanned vehicle， it is urgent to carry out research on multi?category object recognition. In this paper， a multi?category object recognition algorithm based on voxel network is established and its performance is validated. The category label， confidence label and bounding borders regression values of the voxels around the tag are created by calculating the maximal intersection over union（IoU） among prior candidate borders of all categories simultaneously， which resolves the possible mismatch among the three predicted values. The test results indicate that the average recall of category prediction of the proposed multi?category object recognition algorithm is 88.6% and taking the IoU threshold of 0.5 as the correct one， the border regression is 84.8%. Compared with the single?category object recognition network， each category performs an obviously improved accuracy using the proposed algorithm， which proves that the multi?category object recognition algorithm effectively enhances the ability of characteristics learning， and contributes to the improvement of the robustness of the object recognition network.
Aiming at the defects of poor identification effects and prone to be disturbed when using traditional single sensor in vehicle detection， a vehicle detection method based on the fusion of millimeter wave radar and machine vision is propose in this paper. Firstly， the radar data is processed by using hierarchical clustering algorithm with invalid targets filtered out， and the improved YOLO v2 algorithm is adopted to reduce the missed detection rate and increase the detection speed. Then， the intersection?over?union （IoU） of target detection and the global nearest neighbor data association algorithm are utilized to achieve multi?sensor data fusion. Finally， the extended Kalman filter algorithm is employed for target tracking， with the final result obtained. The results of real vehicle test show that the results of vehicle identification with the method proposed is better than that with single sensor， and has good recognition effects under various road conditions.
The adhesion coefficient of water?covered or wet road surfaces is much smaller than that of dry road surfaces， which has a great effect on the traffic safety and maneuverability of vehicle， so by timely obtaining the information of road conditions and issuing warning can greatly reduce potential impairments. In this paper， the application of image?based semantic segmentation network to the recognition of water?covered and wet road conditions is studied， which can not only predict the road conditions in the future， but also obtain the distribution of water?covered and wet areas on road. The method proposed utilizes the semantic segmentation network Res-UNet++ to segment the water?covered and wet areas on road pavement. The Res-UNet++ structure includes an embedded encoder?decoder structure of different depths， with a residual structure added on the feature extraction part of the network to make image features easier to learn. The method adopted achieves an average segmentation accuracy of 90.07% in MIoU， and overcomes the defects of other methods.
Aiming at the current problem of the low accuracy of point cloud segmentation and recognition algorithm in object detection of autonomous vehicle， a sparse color point cloud （SCPC） structure is proposed， which is formed by spatial matching and feature superposition of the image information collected by camera and the point cloud information acquired from lidar. Then， the improved PointPillars neural network algorithm is adopted to conduct operation on the fused SCPC. The results of experiment show that this method can achieve a major rise in average accuracy， compared with original PointPillars algorithm， especially the recognition accuracy of pedestrians and cyclists. The average accuracy of pedestrian and cyclist detections on 3D view under medium difficulty increases by 13.8% and 6.6% respectively， demonstrating the effectiveness of the method adopted.
In view of the threat of abnormal vehicle behavior in V2X， a novel detection method of vehicle abnormal behavior suitable for V2X is proposed in this paper by fusing a variety of machine learning schemes. Firstly， based on Veins V2X simulation platform， various network attacks such as DoS， Sybil， etc. are simulated， the scenes of V2X subject to network attacks under real road conditions are constructed， and the detection data set of abnormal vehicle behavior is built. Then by adopting the idea of stacking ensemble learning and fusing five primary classifiers of K?nearest neighbors， decision tree， multilayer perceptron， AdaBoost， and random forest， an ensemble detection model is set up. Finally， by utilizing the idea of cross?validation， the data set for training is trained by five primary classifiers， with the results of prediction on the data set for validation by primary classifiers as the input of secondary classifier， and the output of secondary classifier as the result of final prediction. The results show that the method proposed has a good detection effect on different network attacks in different scenes of attack density， and a better detection performance than other single classifiers， verifying the effectiveness of the method proposed.
The identification of road roughness is one key technology of smart chassis such as semi?active suspension control. There is a lack of cheap， reliable， accurate and rapid method currently. This paper proposes a new real-time road roughness level identification method based on LSTM （Long Short?Term Memory） network and sequential wheel center acceleration. This method adopts sequential signals of wheel center acceleration instead of traditional statistical features. Based on the LSTM network’s strong feature capture capability for sequential signals， it can rapidly obtain road classification features without signal preprocessing， greatly reducing the calculation burden to realize real?time identification. For training set data， the acceleration signal can be obtained from experimental data， or calculated through vehicle transfer characteristics with white?noise?generated road with different power spectral density levels. This method requires only one time?domain acceleration signal without complex preprocessing. It can achieve rapid identification of road roughness grades at different vehicle speeds， damping coefficients， sprung mass and sampling time with high robustness.
Traditional artificial potential field method has the problems of speed oscillation in longitudinal motion planning， and difficulty in realization of lateral motion planning. A virtual force model based on artificial potential field is established in this paper and a motion planning method of vehicle based on virtual force model of potential field under intelligent connected traffic environment is proposed. By evaluating the motion state of the vehicle and its surrounding vehicles， a virtual force based on vehicle position and speed is generated， and the driving trajectory and speed planning of non?oscillating car following and lane changing are realized. The simulation results show that the proposed method can achieve safe， feasible and smooth collision?free path planning， and can overcome the oscillation problem in traditional potential field motion planning process. By comparing with the real trajectory in driving dataset of highD， it shows that the motion planning based on the virtual force of potential field is approximately consistent with the data of highD， which demonstrates its practicability.
In view of that unreliable network communication brings hidden trouble to V2V control， the design method of connected vehicle controller considering the factors of time delay and packet drop is studied in this paper. Firstly， based on Markovian jump linear system， a model for network control system with stochastic packet drop and time delay events is set up， the stable conditions of the system are proposed in the form of linear matrix inequality， and the design method of the quantized controller with stochastic packet drop is given. On this basis， through the augmentation of system equation matrix， the design method of jump controller under discrete delay is proposed. Finally， a simulation on two typical cloud control scenes of longitudinal and lateral controls of connected vehicle is conducted and the results indicate that under the delay or packet drop with known probability distribution， the quantized controller of connected vehicle control system designed with the method proposed can ensure the stability and safety of system in the situation of unreliable communication.
Considering the problem that the existing trajectory planning algorithms for connected and autonomous vehicle （CAV） passing through intersections cannot achieve optimal coordination of traffic efficiency and safety， the concept of dynamic distance windows （DDW） is introduced according to different initial driving states when CAV drives into the communication range of the intersection， and a trajectory planning algorithm which is suitable for the optimal traffic efficiency under the controllable and safe driving condition is proposed. The DDW corresponding to the initial driving state of CAV can be obtained according to the information of initial driving state parameters of CAV and the signal lights， and the constraints of the maximum comfortable acceleration/deceleration and the speed limit of road. For the two types of situation that the distance between the initial position of CAV and a certain position in front of the stop line is within or outside of the DDW， the corresponding trajectory planning algorithm is designed respectively to achieve the minimum delay of CAV passing through the intersection. The simulation results show that the proposed algorithm can improve the efficiency of CAV passing through the intersection， with smaller speed fluctuations and smoother spatiotemporal trajectory of CAV.
A path planning algorithm based on the modified artificial potential field method and a path tracking optimal guaranteed cost control strategy based on the linear robust control theory for overtaking maneuver of autonomous vehicle on curved road are proposed in this paper. Firstly， the attractive potential field on curve， the repulsive potential field of the leading vehicle with lower speed on the same lane and the repulsive potential fields of curve boundary are built based on the spiral descent function， the slope sine function and the exponential function ，respectively， which form the search space of overtaking path on curve. Then， an incremental search algorithm， which can be applied to dynamic circumstance， is designed to search the fastest descent direction in search space for overtaking path on curve step by step， and the planned overtaking path on curve is obtained. In order to implement the planned path on curve， an overtaking path tracking error dynamic model for curve with parameter perturbation is established based on the 2-DOF linear vehicle dynamics model， and a path tracking optimal guaranteed cost control strategy for overtaking on curve is designed based on the linear robust control theory. Finally， a simulation is performed to verify the feasibility and effectiveness of the path planning algorithm and path tracking optimal guaranteed cost control strategy and the results show that the path planning algorithm and path tracking control strategy proposed can guide the vehicle to complete the overtaking maneuver on curve safely and comfortably.
In view of the fact that most of the existing intelligent vehicle path tracking algorithms take little account of the driver characteristics， an intelligent driving path tracking algorithm based on driver characteristic is proposed. Firstly， k?means algorithm is used to cluster and analyze the relevant data obtained from the real vehicle tests.According to the regularity and difference of handling characteristic parameters， the characteristics of drivers are divided into three types： normal type， radical type and conservative type.Then， according to the classification and clustering results of driver characteristics， the different preferences of different types of drivers for vehicle lateral and longitudinal driving state are integrated into the design of trajectory tracking control strategy. Finally， the intelligent driving path tracking controller is designed based on the model predictive control and the cost function and constraints of the controller are designed based on the results of data clustering. The simulation results demonstrate that the vehicle path tracking control strategy proposed in this paper has high tracking accuracy and speed control accuracy. And the vehicle response changes in the tracking process can reflect the characteristics of different drivers. The path tracking speed error is less than 2% and the lateral tracking error is less than 0.13 m.
The motion characteristics of the autonomous vehicle should be taken into account to ensure safety， comfort and stability when avoiding obstacles. An obstacle avoidance trajectory planning algorithm with the constraints of vehicle motion is proposed in this paper. Firstly， combining with the motion state and position information of obstacles， the algorithm derives the derivative status region where the vehicle can safely avoid obstacles from the local environment map. Then， the terminal state points are sampled in the derivative status region to form a discrete terminal state point set. The obstacle avoidance trajectory search problem in complex road environment is transformed into the problem of trajectory fitting and optimization between the vehicle and the state point set. Trajectory fitting is realized by Bézier curve planner based on vehicle lateral dynamics model while the optimization process takes into consideration of driving smoothness and operation stability in the process of vehicle trajectory following. By comparing the obstacle avoidance effect in multiple scenarios with the conventional State Lattice algorithm and MPC algorithm， the results show that the proposed algorithm can make the vehicle avoid obstacles safely and reasonably in the test scenarios， and has better performance in the aspects of trajectory smoothness and vehicle handling stability.
To enhance the longitudinal car?following performance of intelligent driving system， a driver's car?following model based on deep reinforcement learning is constructed in this paper. Firstly， according to the selection rule defined of car following scenes， typical car?following scenes conforming to conditions are selected from the natural driving data， on which a statistical analysis is then conducted to analyze the influence mechanism of the factors of car spacing， relative speed and time headway on the car following behavior of driver by using correlation coefficient method， with the behavior characteristic and its affecting factors of driver's car following driving process obtained. Then a car following model of driver is established based on the deep deterministic policy gradient algorithm， and the driver's data set of car following trajectory is input into the simulated car following environment so that the intelligent agent can learn the decision?making behavior of driver from the empirical data. Finally， with the original data as the reference base， a comparative simulation verification is performed on the deep reinforcement learning?based car following model， with a result showing that the driver's car following model constructed has good tracking performance and can truly reproducing the car following behavior of driver.
To enhance the estimation accuracy of key parameters and reduce the effects of model uncertainty on the robustness of control system in the autonomous path tracking process of distributed?drive intelligent electric vehicle， an observer?based adaptive sliding mode path tracking control strategy is proposed in this paper. Firstly， in view of the difficulty in directly and accurately measuring the longitudinal and lateral speeds， a state estimation system with 5 inputs， 3 outputs and 3 states is established， and the minimal model error criterion is adopted to reduce the error of observation model caused by the nonlinear feature of tire. Then based on kinematic model， the desired yaw rate response of path tracking is calculated， the sliding mode algorithm is employed to achieve active steering control， and with consideration of the potential failure risk of steer?by?wire system， the RBF neural network is introduced to perform an online estimation on system uncertainty. Meanwhile， the direct yaw controller is designed with optimal torque distribution strategy used to further improve the stability of vehicle. Finally， a Carsim/Matlab co?simulation is conducted on vehicle state estimation and path tracking. The results demonstrate that the observer based on minimal model error criterion can get more reliable estimation results and the path tracking controller can ensure the vehicle have higher tracking accuracy and robustness.
Aiming at the problem of collision avoidance in the process of vehicle longitudinal following， an improved safety distance model is proposed based on the factors of vehicle motion state and road adhesion coefficient. In view of the issue of coordinated control between driver and active braking system， the method of extension decision is adopted to establish a 2D extension set， with the real vehicle spacing and time?to?collision as reference variables. The dynamic safety boundary is divided， and free driving mode， coordinated braking mode and active braking mode are adopted respectively in different domains. Based on collision avoidance model， the model for radial basis function neural network selected as active braking controller is trained to obtain the ideal braking pressure. Both software simulation and the hardware （test bench）?in?the?loop simulation are conducted to verify the control strategy proposed， and the results show that the strategy proposed can effectively avoid the vehicle longitudinal collision， and improve the braking smoothness and safety.
In order to study the relationship between the driver’s inhibitory control ability and risky driving behavior， the cued Go/NoGo experimental paradigm is used to measure the driver’s inhibitory control ability， a simulated driving scene is built by combining the characteristics of the inhibitory control and the risk decision behavior. Based on the driving simulator system， 51 testees take part in the simulated driving experiments， the performances of drivers’ risk decision?making are recorded， the driver’s operation and vehicle operation data are collected， and the effects of inhibitory control ability on risky driving behavior are analyzed. The results show that the drivers with low inhibitory control ability often show higher risk behavior tendency when facing risk decision?making and tend to take higher driving speed with a higher proportion of overspeed behavior and a weaker ability of vehicle lateral control.
For autonomous vehicle test scenarios construction， this paper firstly makes a comparative analysis of the existing scenario definition and architecture， and proposes that test scenarios should cover a total of 10 layers of information of scene elements and test elements. Secondly， a method system of scenario construction including direct construction of concrete scenario， mining and deduction of typical logical scenario and reconstruction and derivation of specific scenario is summarized and proposed. Thirdly， from the three dimensions of single segment test， combined segment test and integrated traffic flow test， the main virtual scenario test application methods are systematically sorted out. Finally， the research prospect is presented from the perspective of method chain and tool chain of scenario construction. The research results of the review will provide reference for the testing and evaluation of autonomous vehicles.
In order to solve the problems of uncertain road test scenes for autonomous vehicles， lack of typical evaluation scenes， and the difficulty in quantitatively evaluating scenes， a road test evaluation method for autonomous vehicles based on scene complexity model is proposed， which evaluate and classify the road test scenes for autonomous vehicles by innovatively combining information entropy with gravity model， and the rationality of the scene evaluation method is verified by tests. The method is helpful for the autonomous driving enterprises and the third-party testing / evaluation organizations to select the typical scenes， promoting the course of road test and evaluation for autonomous vehicles.
The attention state of driver is an important human factor affecting traffic safety， and spatial attention orientation is a key subfunction of attention. In order to explore the relationship between the dangerous driving behavior and the spatial attention orientation function of driver， a total of 37 healthy testees are invited to participate in simulated driving experiment and cue?target stimulation experiment. The testees are divided into three groups according to the results of cluster analysis on their dangerous driving behaviors. Then the behavioral performance and electroencephalogram data of testees in three groups in cue?target stimulation experiment are compared， and a statistical analysis on the effectiveness of cue and the results of cue?stimulus interval is conducted by using paired?T test and the variance analysis of repeated measures. The results of analysis on the reaction time and error rate of testees show that the testees having lower error rate are higher in behavior performance and are weaker in cuing effects， well relating to their reasonable distribution of attention resources and better ability in expressing external stimulus and own emotions.
