To clarify the limitations and development direction of the research on Collaborative Intersection Collision Warning (CICW), CICW research progress was reviewed systematically. Firstly, the advantages and disadvantages of the existing intersection traffic conflict detection methods were analyzed. Secondly, the applicability of different warning levels, mechanisms, and modes was summarized. Thirdly, the effectiveness and safety evaluation indexes of CICW were identified, and a comparison was made between simulation, real vehicle and virtual-real fusion testing. Then, the influence of driver uncertainty and unreliable communication on CICW and its optimization were analyzed. Finally, the future development direction of CICW was prospected. The results show that the traffic risk field model presents a feasible solution to solve the comprehensive characterization of risk and conflict severity in the existing CICW conflict detection methods. However, further research is needed to establish appropriate environmental parameters, risk indicators, and determination. Moreover, modeling, forecasting, and online identification of driver’s behavior in CICW application scenarios and the adaptive construction of CICW warning modes offer viable solutions to designing reliable and effective CICW applications. Achieving a comprehensive objective evaluation of CICW in all aspects depends on the research and establishment of a comprehensive evaluation mechanism and a large-scale experimental platform. The unreliable Internet Of Vehicle (IOV) communication seriously affects the effectiveness of CICW, so it is necessary to further study the channel congestion control mechanism and CICW fault-tolerant mechanism based on communication failure/failure prediction.

In order to further improve the fuel economy of autonomous vehicle driving at intersections, this paper analyzed quantitatively the performance index functions and constraints including vehicle safety, economy, comfort, etc., based on the Model Predictive Control (MPC) theory, and an ecological driving controller of autonomous vehicle at intersections was designed. The simulation results show that the control strategy proposed in the paper can ensure good safety and comfort, reduce fuel consumption by 15.83% and 34.98% respectively compared with Linear Quadratic Regulator (LQR) under conditions with/without vehicle ahead affecting vehicle driving.

In current research on vehicle trajectory prediction, the existing Graph Attention Network (GAT), which is based on a static attention mechanism, fails to effectively capture interactions between vehicles in complex road conditions. To address this issue, this paper proposed an Encoder-Decoder Dynamic Graph Attention Network (ED-DGAT) to predict future trajectories of highway vehicles. In this model, the encoding module incorporates a dynamic graph attention mechanism to learn spatial interactions among vehicles. Simultaneously, a simplified dynamic graph attention network is adopted to model the interdependencies of vehicle movements during the decoding phase. This paper evaluated the proposed algorithm using the NGSIM dataset and conducted comparative analysis with other models such as LSTM, Social-LSTM (S-LSTM), and CS-LSTM. The results show that the Root Mean Squared Error (RMSE) of predicted trajectory has been reduced by 25%, and the inference speed is 2.61 times of the CS-LSTM model.

In order to improve the shortcomings of poor smoothness and potential collision in traditional Rapidly-exploring Random Tree (RRT) algorithm for global path planning, the paper proposed a dual-optimization RRT algorithm. Based on the traditional RRT algorithm, an adaptive target bias strategy was introduced to shorten the sampling time, and an angle-constrained sampling strategy was introduced to adapt to the vehicle’s maximum steering angle. After the initial path was obtained, a binary optimization function (reducing path curvature and avoiding obstacles) was established and used as a basis for gradient descent secondary optimization, generating a path that can be driven by vehicles with good smoothness and low collision probability, which was then simulated and verified. The results show that compared with RRT algorithm, RRT-Connect algorithm and RRT* algorithm, the optimized RRT algorithm reduces average curvature by 38.1%, 36.4% and 24.7%, respectively; while reducing curvature variance by 38.4%, 38.4% and 27.2%, respectively.

For the fact that single control algorithm cannot simultaneously meet the requirements of autonomous vehicles for path tracking accuracy and controller solving speed, this paper proposed a hybrid control strategy based on Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC). The strategy used an LQR in the low-speed condition and an MPC algorithm in the high-speed condition, on the basis of which a switching mechanism of the control algorithm based on a Finite State Machine (FSM) was designed and the control parameters were optimized by Genetic Algorithm (GA). The hybrid control strategy was simulated and verified based on CarSim and MATLAB/Simulink simulation platforms, and the real vehicle test was further completed. The experimental results show that the designed hybrid control strategy can reduce the computation time on the basis of improving the tracking accuracy, and the average lateral error and average heading error are reduced by 26.3% and 39.6%, respectively, and the average computation time is reduced by 10.9% compared with the single control algorithm.

In order to solve the problem of low control accuracy and low parameter tuning efficiency caused by difficulty in selecting coefficient matrix Q and R of Linear Quadratic Regulator (LQR) in lateral control of intelligent vehicle, this paper proposed an optimization method of genetic particle mixing (Genetic Algorithm -Particle Swarm Optimization, GA-PSO). A lateral LQR controller and a feed-forward controller were designed based on the two-degree-of-freedom model of the vehicle. The coefficient matrix was optimized using the LQR controller’s own energy loss function as the cost function. The algorithm optimization results of GA-PSO and PSO were compared. The CarSim/Simulink co-simulation shows that the GA-PSO optimized controller improves the tracking accuracy and computing efficiency by 47.06% and 63.54%, respectively, and the optimized controller has strong robustness.

Based on the historical driving data of trucks in a province, this paper proposed a prediction method of dangerous driving behavior based on Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) network and self-attention mechanism. For the characteristics of large amount of truck driving data, high dimension, difficult feature extraction and strong time sequence, this method first used XGBoost to filter the features, then used CNN to extract spatial features and LSTM to further capture the temporal information of driving behaviors. Finally, dangerous driving behaviors were predicted by self-attention mechanism. Experimental results show that the method presented in this paper performs better than other long time series prediction methods on highway freight driving data in a province, with recognition accuracy reaching 85.05%, the weighted average recall rate reaches 83%, and the F1-score reaches 84%.