Search Results

The on-ramp merging driving scenario is challenging for achieving the highest-level autonomous driving. Current research using reinforcement learning methods to address the on-ramp merging problem of automated vehicles (AVs) is mainly designed for a single AV, treating other vehicles as part of the environment. This paper proposes a control framework for cooperative on-ramp merging of multiple AVs based on multi-agent deep reinforcement learning (MADRL). This framework facilitates AVs on the ramp and adjacent mainline to learn a coordinate control policy for their longitudinal and lateral motions based on the environment observations. Unlike the hierarchical architecture, this paper integrates decision and control into a unified optimal control problem to solve an on-ramp merging strategy through MADRL.

In the field of vehicle aerodynamic simulation, Reynold Averaged Navier-Stokes (RANS) model is widely used due to its high efficiency. However, it has some limitations in capturing complex flow features and simulating large separated flows. In order to improve the computational accuracy within a suitable cost, the Field Inversion and Machine Learning (FIML) method, based on a data-driven approach, has received increasing attention in recent years. In this paper, the optimal coefficients of the Generalized k-ω (GEKO) model are firstly obtained by the discrete adjoint method of FIML, utilizing the results of wind tunnel experiments. Then, the mapping relationship between the flow field characteristics and the optimal coefficients is established by a neural network to augment the turbulence model.

Reinforcement learning is a promising approach to solve the energy management for hybrid electric vehicles. In this paper, based on the DQN (Deep Q-Network) reinforcement learning algorithm which is widely used at present, double DQN, dueling DQN and learning from demonstration are integrated; states, actions, rewards and the experience pool based on the characteristics of series-parallel multi-speed hybrid powertrain are designed; the hybrid energy management strategy based on D4QN (Double Dueling Deep Q-Network with Demonstrations) algorithm is established. Based on the training results of D4QN algorithm, multi-parameter analysis under state and action space, HCU (Hybrid control unit) application and MIL (Model in-loop) test research are conducted.

The driver monitoring system (DMS) plays an essential role in reducing traffic accidents caused by human errors due to driver distraction and fatigue. The vision-based DMS has been the most widely used because of its advantages of non-contact and high recognition accuracy. However, the traditional RGB camera-based DMS has poor recognition accuracy under complex lighting conditions, while the IR-based DMS has a high cost. In order to improve the recognition accuracy of conventional RGB camera-based DMS under complicated illumination conditions, this paper proposes a lightweight low-illumination image enhancement network inspired by the Retinex theory. The lightweight aspect of the network structure is realized by introducing a pixel-wise adjustment function. In addition, the optimization bottleneck problem is solved by introducing the shortcut mechanism.

Harmony-with-traffic refers to the ability of autonomous vehicles to maximize the driving benefits such as comfort, efficiency, and energy consumption of themselves and the surrounding traffic during interactive driving under traffic rules. In the test of harmony-with-traffic, one or more background vehicles that can respond to the driving behavior of the vehicle under test are required. For this purpose, the functional requirements of car-following model for harmony-with-traffic evaluation are analyzed from the dimensions of test conditions, constraints, steady state and dynamic response. Based on them, an interactive car-following model (ICFM) is developed. In this model, the concept of equivalent distance is proposed to transfer lateral influence to longitudinal. The calculation methods of expected speed are designed according to the different car-following modes divided by interaction object, reaction distance and equivalent distance.

The intersection is recognized as the most dangerous area because of the restricted road structures and indeterminate traffic regulations. Therefore, according to the Vehicle-to-everything (V2X) communication, Intelligent Transportation Systems (ITS), and Digital Twin data, we present a potential energy field method to establish the general characteristics of intersection traffic safety, evaluate the safety situation of intersection and assist intersection traffic participants in passing through the intersection safer and more efficient. The resulting potential energy field method is established by the contour line of traffic participants' potential energy, which is constructed as a superposition of disparate energies, such as boundary potential energy, body potential energy, and velocity potential energy. The intersection traffic safety is evaluated by the potential energy field characteristic of simultaneous intersection traffic participants.

With the popularity of automated vehicles, the future mixed traffic flow contains automated vehicles with different degrees of intelligence developed by other manufacturers. Therefore, simulating the interaction behavior of automated vehicles with varying levels of intelligence is crucial for testing and evaluating autonomous driving systems. Since the algorithm of traffic vehicles with various intelligence levels is difficult to obtain, it leads to hardships in quantitatively characterizing their interaction behaviors. Therefore, this paper designs a new automated vehicle test platform to solve the problem. The intelligent vehicle testbed with multiple personalized in-vehicle control units in the loop consists of three parts: 1. Multiple controllers in the loop to simulate the behavior of traffic vehicles；2. The central console applies digital twin technology to share the same traffic scenario between the tested vehicle and the traffic vehicle, creating a mixed traffic flow. 3.

High-speed vehicles in low illumination environments severely blur the images used in object detectors, which poses a potential threat to object detector-based advanced driver assistance systems (ADAS) and autonomous driving systems. Augmenting the training images for object detectors is an efficient way to mitigate the threat from motion blur. However, little attention has been paid to the motion of the vehicle and the position of objects in the traffic scene, which limits the consistence between the resulting augmented images and traffic scenes. In this paper, we present a vehicle kinematics-based image augmentation algorithm by modeling and analyzing the traffic scenes to generate more realistic augmented images and achieve higher robustness improvement on object detectors against motion blur. Firstly, we propose a traffic scene model considering vehicle motion and the relationship between the vehicle and the object in the traffic scene.

The research and development of data-driven highly automated driving system components such as trajectory prediction, motion planning, driving test scenario generation, and safety validation all require large amounts of naturalistic vehicle trajectory data. Therefore, a variety of data collection methods have emerged to meet the growing demand. Among these, camera-equipped drones are gaining more and more attention because of their obvious advantages. Specifically, compared to others, drones have a wider field of bird's eye view, which is less likely to be blocked, and they could collect more complete and natural vehicle trajectory data. Besides, they are not easily observed by traffic participants and ensure that the human driver behavior data collected is realistic and natural. In this paper, we present a complete vehicle trajectory data extraction framework based on aerial videos. It consists of three parts: 1) objects detection, 2) data association, and 3) data cleaning.

A formal gesture recognition based on optics has limitations, but millimeter-wave (MMW) radar has shown significant advantages in gesture recognition. Therefore, the MMW radar has become the most promising human-computer interaction equipment, which can be used for human-computer interaction of vehicle personnel. This paper proposes a multi-branch network based on a residual neural network (ResNet) to solve the problems of insufficient feature extraction and fusion of the MMW radar and immense algorithm complexity. By constructing the gesture sample library of six gestures, the MMW radar signal is processed and coupled to establish the relationship between the motion parameters of the distance, speed, and angle of the gesture information and time, and the depth features are extracted. Then the three depth features are fused. Finally, the classification and recognition of MMW radar gesture signals are realized through the full connection layer.

Deep neural network models have been widely used for environment perception of intelligent vehicles. However, due to models’ innate probabilistic property, the lack of transparency, and sensitivity to data, perception results have inevitable uncertainties. To compensate for the weakness of probabilistic models, many pieces of research have been proposed to analyze and quantify such uncertainties. For safety-critical intelligent vehicles, the uncertainty analysis of data and models for environment perception is especially important. Uncertainty estimation can be a way to quantify the risk of environment perception. In this regard, it is essential to deliver a comprehensive survey. This work presents a comprehensive overview of uncertainty estimation in deep neural networks for environment perception of intelligent vehicles.

For high levels autonomous driving functions, the Decision Layer often takes on more responsibility due to the requirement of facing more diverse and even rare conditions. It is very difficult to accurately find a safe and efficient lane change timing when autonomous vehicles encounter complex traffic flow and need to change lanes. The traditional method based on rules and experiences has the limitation that it is difficult to be taken into account all possible conditions. Therefore, this paper designs a lane-changing decision algorithm based on data-driven and machine learning, and uses the DQN (Deep Q Network) algorithm in Reinforcement Learning to determine the appropriate lane-changing timing and target lane. Firstly, the scene characteristics of the highway are analyzed, the input and output of the decision-making model are designated and the data from the Perception Layer are processed.

Based on the sample data obtained from the bench test of a four-cylinder naturally aspirated CNG engine, three different machine learning models, BP, SVM and GRNN, were used to develop the intake charge prediction model for the intake system of this engine, in which engine speed, intake manifold pressure and intake temperature, VVT angle and gas injection time were taken as input parameters and intake charge was used as output parameter. The comparative analysis of the experimental data and model prediction data showed that the mean absolute error (MAE) of BP model, GRNN model, and SVM model were 2.69, 8.11and 5.13, and the root mean square error (MSE) were 3.53, 9.29, and 7.17, respectively. BP model has smaller prediction error and higher accuracy than SVM and GRNN models, which is more suitable for the prediction of the intake charge of this type of four-cylinder naturally aspirated CNG engine.

As turbulence modeling has become an indispensable approach to perform flow simulation in a wide range of industrial applications, how to enhance the prediction accuracy has gained increasing attention during the past years. Of all the turbulence models, RANS is the most common choice for many OEMs due to its short turn-around time and strong robustness. However, the default setting of RANS is usually benchmarked through classical and well-studied engineering examples, not always suitable for resolving complex flows in specific circumstances. Many previous researches have suggested a small tuning in turbulence model coefficients could achieve higher accuracy on a variety of flow scenarios. Instead of adjusting parameters by trial and error from experience, this paper introduced a new data-driven method of turbulence model recalibration using adjoint solver, based on Generalized k-ω (GEKO) model, one variant of RANS.

Motivated by applying artificial intelligence technology to the automobile industry, reinforcement learning is becoming more and more popular in the community of intelligent driving research. The reward function is one of the critical factors which affecting reinforcement learning. Its design principle is highly dependent on the features of the agent. The agent studied in this paper can do perception, decision-making, and motion-control, which aims to be the assistant or substitute for human driving in the latest future. Therefore, this paper analyzes the characteristics of excellent human driving behavior based on the six-layer model of driving scenarios and constructs it into a human knowledge graph. Furthermore, for highway pilot driving, the expert demo data is created, and the reward function is self-learned via inverse reinforcement learning. The reward function design method proposed in this paper has been verified in the Unity ML-Agent environment.

Object detection is an important visual content of the autonomous vehicle, the traditional detecting methods usually cost a lot of computational memory and elapsed time. This paper proposes to use lightweight deep convolutional neural network (MobilenetV3-SSDLite) to carry out the object detection task of autonomous vehicles. Simulation analysis based on this method is implemented, the feature layer obtained after h-swish activation function in the first Conv of the 13th bottleneck module in MobilenetV3 is taken as the first effective feature layer, and the feature layer before pooling and convolution of the antepenultimate layer in MobilenetV3 is taken as the second effective feature layer, and these two feature layers are extracted from the MobilenetV3 network.

The main challenge for future autonomous vehicles is to identify their location and body pose in real time during driving, that is, “where am I? and how will I go?”. We address the problems of pose estimation and scene perception from continuous visual frames in intelligent vehicle. Recent advanced technology in the domain of deep learning proposes to train some learning models for vehicle’s series detection tasks in a supervised or unsupervised manner, which has numerous advances over traditional approaches, mainly reflected in the absence of manual calibration and synchronization of the camera and IMU. In the paper, we propose a novel approach for pose estimation and scene recognition with a deep fusion of multi-modal neural features in the manner of unsupervised. Firstly, low-cost camera and IMU are used to extract original visual and inertial data, then the visual and inertial encoders are utilized to encoder the feature of the two modes.

The plug-in hybrid electric vehicles with dual-motor and multi-gear structure can realize multiple operation modes such as series, parallel, hybrid, etc. The traditional rule-based energy management strategy mostly selects some of the modes (such as series and parallel) to construct the energy management strategy. Although this method is simple and reliable, it can’t fully exert the full potential of this structure considering both economy and driving performance. Therefore, it is very important to study the algorithm which can exert the maximum potential of the multi-degree-of-freedom structure. In this paper, a new RL-PMP algorithm is proposed, which does not divide the operation modes, and explores the optimal energy allocation strategy to the maximum extent according to the economic and drivability criteria within the allowable range of the characteristics of the power system components.

In the autonomous driving perception task, the accuracy of target detection is an essential evaluation, especially for small targets. In this work, we propose a multi-sensor fusion neural network that combines radar and image data to improve the confidence level of the camera when detecting targets and the accuracy of the prediction box regression. The fusion network is based on the basic structure of single-shot multi-box detection (SSD). Inspired by the attention mechanism in image processing, our work incorporates the a priori knowledge of radar detection in the convolutional block attention module (CBAM), which forms a new attention mechanism module called radar convolutional block attention module (RCBAM). We add the RCBAM into the SSD target detection network to build a deep neural network fusing millimeter-wave radar and camera.

Since situational awareness is crucial for autonomous driving in urban environments, multi-target tracking has become an increasingly popular research topic during the last several years. For autonomous driving in urban environments, cars and pedestrians are the two main types of obstacles, and their motion characteristics are not the same. While in the current related multi-target tracking research, the same motion model (such as Constant Velocity model [CV]) or motion model set (such as CV combined with Constant Acceleration model [CA]) is mostly used to track different types of obstacles simultaneously. Besides, in current research, regular motion models are mostly adopted to track pedestrians, such as CV, CA, and so on, the uncertainty in pedestrian motion is not well considered.