Search Results

Forecasting the motion of the leading vehicle is a critical task for connected autonomous vehicles as it provides an efficient way to model the leading-following vehicle behavior and analyze the interactions. In this study, a personalized time-series modeling approach for leading vehicle trajectory prediction considering different driving styles is proposed. The method enables a precise, personalized trajectory prediction for leading vehicles with limited inter-vehicle communication signals, such as vehicle speed, acceleration, space headway, and time headway of the front vehicles. Based on the learning nature of human beings that a human always tries to solve problems based on grouping and similar experience, three different driving styles are first recognized based on an unsupervised clustering with a Gaussian Mixture Model (GMM).

To reduce the interference and conflict of human-machine cooperative control, lighten the operation workload of drivers, and improve the friendliness and acceptability of intelligent vehicles, a personalized human-machine cooperative lane-change trajectory tracking control method was proposed. First, a lane-changing driving data acquisition test was carried out to collect different driving behaviors of different drivers and form the data pool for the machine learning method. Two typical driving behaviors from an aggressive driver and a moderate driver are selected to be studied. Then, a control structure combined by feedforward and feedback control based on Long Short Term Memory (LSTM) and model-based optimum control was introduced. LSTM is a machine learning method that has the ability of memory. It is used to capture the lane-changing behaviors of each driver to achieve personalization. For each driver, a specific personalized controller is trained using his driving data.

Autonomous driving technology, as the product of the fifth stage of the information technology revolution, is of great significance for improving urban traffic and environmentally friendly sustainable development. Autonomous driving can be divided into three main modules. The input of the decision module is the perception information from the perception module and the output of the control strategy to the control module. The deep reinforcement learning method proposes an end-to-end decision-making system design scheme. This paper adopts the Deep Deterministic Policy Gradient Algorithm (DDPG) that incorporates the Priority Experience Playback (PER) method. The framework of the algorithm is based on the actor-critic network structure model. The model takes the continuously acquired perception information as input and the continuous control of the vehicle as output.

A feasible method was developed to reconstruct the three-dimensional flame surface of SI combustion based on 2D images. A double-window constant volume vessel was designed to simultaneously obtain the side and bottom images of the flame. The flame front was reconstructed based on 2D images with a slicing model, in which the flame characteristics were derived by slicing flame contour modeling and flame-piston collision area analysis. The flame irregularity and anisotropy were also analyzed. Two different principles were used to build the slicing model, the ellipse hypothesis modeling and deep learning modeling, in which the ellipse hypothesis modeling was applied to reconstruct the flame in the optical SI engine. And the reconstruction results were analyzed and discussed. The reconstruction results show that part of the wrinkled and folded structure of the flame front in SI engines can be revealed based on the bottom view image.

Accurate identification of driver’s braking intention is essential in advanced driver assistance system and can make the driving process more comfortable and trustworthy. In this paper, a novel method for driver braking intention identification in cut-in scenarios was proposed by using driver’s gaze information and motion information of cut-in vehicles. Firstly, a "looking in and looking out" experimental platform including three eye-tracking cameras and one front-view camera was built to collect driver's gaze information and the vehicle motion information. Secondly, driver’s gaze features and motion features of cut-in vehicles were selected and the braking intention identification performance of several decision tree-based ensemble learning algorithms was compared. Thirdly, the feature importance was analyzed by using SHAP (SHapley Additive exPlanations) values. This novel method of braking intention identification makes full use of in-vehicle camera sensors.

Driving style and skill classification are of great significance in human-oriented advanced driver-assistance system (ADAS) development. In this paper, we propose Fully-Connected Deep Neural Networks (FC-DNN) to classify drivers’ styles and skills with naturalistic driving data. Followed by the data collection and pre-processing, FC-DNN with a series of deep learning optimization algorithms are applied. In the experimental part, the proposed model is validated and compared with other commonly used supervised learning methods including the k-nearest neighbors (KNN), support vector machine (SVM), decision tree (DT), random forest (RF), and multilayer perceptron (MLP). The results show that the proposed model has a higher Macro F1 score than other methods. In addition, we discussed the effect of different time window sizes on experimental results. The results show that the driving information of 1s can improve the final evaluation score of the model.

It’s hard to achieve complete self-driving using hand-crafting generalized decision-making rules, while the end-to-end self-driving system is low in complexity, does not require hand-crafting rules, and can deal with complex situations. Modular-based self-driving systems require multi-task fusion and high-precision maps, resulting in high system complexity and increased costs. In end-to-end self-driving, we usually only use camera to obtain scene status information, so image processing is very important. Numerous deep learning applications benefit from multi-task learning, as the multi-task learning can accelerate model training and improve accuracy with combine all tasks into one model, which reduces the amount of calculation and allows these systems to run in real-time. Therefore, the approach of obtaining rich scene state information based on multi-task learning is very attractive. In this paper, we propose an approach to multi-task learning for semantics, geometry and motion.

According to the statistics of National Highway Traffic Safety Administration, driver’s cognitive distraction, which is usually caused by drivers using mobile phones, has become one of the main causes of traffic accidents. To solve this problem and guarantee the safety of man-vehicle-road system, the most critical work is to improve the accuracy of driver’s cognitive state detection. In this paper, a novel driver’s cognitive state detecting method based on LightGBM (Light Gradient Boosting Machine) is proposed. Firstly, cognitive distraction experiments of making calls are carried out on a driving simulator to collect vehicle states, eye tracking and EEG (electron encephalogram) data simultaneously and feature extraction is conducted. Then a classifier considering road and individual characteristics used for detecting cognitive states is trained based on LightGBM algorithm, with 3 predefined cognitive states including concentration, ordinary distraction and extreme distraction.

Aiming at the problems of long lane change intention recognition, complicated lane change model, and huge amount of processing data in the current research, this paper uses principal component analysis to improve the driver’s lane change intention recognition model using traditional pattern recognition. Firstly collect 7 parameters including driver operation and vehicle running characteristics. After data standardization and PCA (principal component analysis), the top three principal components that can reflect the information content of the original data are nearly 90%. Then, a lane-change intent recognition model based on GMM-HMM was established, three lane change intents cannot be directly observed as the hidden state of the model; and three principal component quantities obtained through linear changes are used as observational measurements.

This work presents a multi-objective adaptive cruise control (ACC) system via deep reinforcement learning (DRL). During the control period, it quantitatively considers three indexes: tracking accuracy, riding comfort, and fuel economy. The system balances contradictions between different indexes to achieve the best overall control results. First, a hierarchical control architecture is utilized, where the upper level controller is synthesized under DRL framework to give out the vehicle desired acceleration. The lower level controller executes the command and compensates vehicle dynamics. Then, four state variables that can comprehensively determine the car-following states are selected for better convergence. Multi-objective reward function is quantitatively designed referring to the evaluation indexes, in which safety constraints are considered by adding violation penalty. Thereafter, the training environment which excludes the disturbance of preceding car acceleration is built.

Batteries are exposed to dynamic load during vehicle driving. It is significant to clarify the load input of the battery system during vehicle driving for battery pack structural design and optimization. Currently, bolt connection is mostly applied for battery pack constraint to vehicle, as well as for module assembly inside the pack. However, accurate bolt load is always difficult to obtain, while directly force measurement is expensive and time consuming in engineering. In this paper, a precise data driven model based on Elman neural network is established to identify the dynamic bolt loads of the battery pack, using tested acceleration data near bolts. The dynamic bolt force data is measured at the same time with the acceleration data during vehicle running in different driving conditions, utilizing customized bolt force sensors.

Roundabouts are one of the most complex traffic environments in urban roads, and a key challenge for intelligent driving decision-making. Deep reinforcement learning, as an emerging solution for intelligent driving decisions, has the advantage of avoiding complex algorithm design and sustainable iteration. For the decision difficulty in roundabout scenarios, this paper proposes an artificial potential field based Soft Actor-Critic (APF-SAC) algorithm. Firstly, based on the Carla simulator and Gym framework, a reinforcement learning simulation system for roundabout driving is built. Secondly, to reduce reinforcement learning exploration difficulty, global path planning and path smoothing algorithms are designed to generate and optimize the path to guide the agent.

Under complex and extreme operating conditions, the road adhesion coefficient emerges as a critical state parameter for tire force analysis and vehicle dynamics control. In contrast to model-based estimation methods, intelligent tire technology enables the real-time feedback of tire-road interaction information to the vehicle control system. This paper proposes an approach that integrates intelligent tire systems with machine learning to acquire precise road adhesion coefficients for vehicles. Firstly, taking into account the driving conditions, sensor selection is conducted to develop an intelligent tire hardware acquisition system based on MEMS (Micro-Electro-Mechanical Systems) three-axis acceleration sensors, utilizing a simplified hardware structure and wireless transmission mode. Secondly, through the collection of real vehicle experiment data on different road surfaces, a dataset is gathered for machine learning training.

With the capability of avoiding failure in advance, failure prediction model is important not only to end users, but also to the service engineers in vehicle industry. This paper proposes an approach based on anomaly detection algorithms and telematic data to predict the failure of the engine air system with Turbo charger. Firstly, the relationship between air system and all obtained features are analyzed by both physical mechanism and data-wise. Then, the features including altitude, air temperature, engine output power, and charger pressure are selected as the input of the model, with the sampling interval of 1 minute. Based on the selected features, the healthy state for each vehicle is defined by the model as benchmark. Finally, the ‘Medium surface’ is determined for specific vehicle, which is a hyperplane with the medium points of the healthy state located at, to detect the minor weakness symptom (sub-health state).

With the rapid advancement in intelligent vehicle technologies, comprehensive environmental perception has become crucial for achieving higher levels of autonomous driving. Among various perception tasks, monitoring road types and evenness is particularly significant. Different road categories imply varied surface adhesion coefficients, and the evenness of the road reflects distinct physical properties of the road surface. This paper introduces a two-stage road perception framework. Initially, the framework undergoes pre-training on a large annotated drivable area dataset, acquiring a set of pre-trained parameters with robust generalization capabilities, thereby endowing the model with the ability to locate road areas in complex regions.