This paper systematically reviews the developmental trajectory of autonomous driving systems from the 1920s to 2024, analyzing four key phases's characteristics—technological origins and infancy, theoretical foundations and exploration, practical breakthroughs, and rapid development with paradigm shifts. It focuses on the modular "perception-decision-control" driving, the end-to-end driving and the synergy and competition within the vehicle-road-cloud collaborative. Research indicates that the current autonomous driving landscape exhibits a "dual-track development pattern dominated by modular driving yet driven by end-to-end driving breakthroughs", where data closed-loops and vehicle-road-cloud integration serve as critical pathways to address long-tail scenarios and enhance system generalization. However, insufficient data quality, lack of algorithmic interpretability, surging computational demands and policy adaptation delays remain the main challenges. Predictions suggest that advancements in large model-powered cognitive reasoning, neuro-symbolic hybrid architectures, lightweight deployment, and deep vehicle-road-cloud collaboration will propel autonomous driving toward a safer, more efficient, and inclusive intelligent mobility ecosystem. This study provides a systematic reference for understanding the technological landscape, optimizing development strategies, and formulating industry policies.

To address the issue that existing point cloud-based 3D object detection algorithms perform poorly on small object detection, a PointPillars 3D object detection algorithm based on an attention mechanism is proposed on the basis of the traditional PointPillars model. Firstly, the pillar feature network in the PointPillars model is improved by introducing a set of Efficient Channel Attention (ECA) mechanisms in the encoding network, allowing each Pillar module to better consider its local detailed geometric information, thereby significantly enhancing the feature representation capability of each Pillar module. Secondly, a Shuffle Attention module is introduced into the 2D convolutional downsampling module of the backbone network. By improving the diversity and effectiveness of feature representation, the model network focuses on important features while suppressing less important or redundant features, significantly enhancing the algorithm's feature extraction ability. Test results on the KITTI dataset show that the improved algorithm improves the 3D pedestrian detection accuracy by 4.04 percentage points, 3.43 percentage points and 1.93 percentage points for easy, moderate, and hard difficulties, respectively, significantly outperforming the original PointPillars network and enhancing the performance of small pedestrian object detection.

To address the limitations of traditional lane-keeping methods that rely on high-fidelity models and significant computational burden, this paper proposes a lane-keeping assist strategy based on the Deep Deterministic Policy Gradient (DDPG) algorithm. The proposed system integrates vehicle dynamics models with a Deep Reinforcement Learning (DRL) framework. Furthermore, a specialized reward function considering both task completion and lateral error is designed to optimize training convergence speed and tracking accuracy, thereby achieving more precise and efficient performance across various curve scenarios. Simulation experiments demonstrate that the proposed strategy yields significantly smaller lateral errors compared to the traditional Model Predictive Control (MPC) method. Specifically, the Mean Absolute Lateral Error (MALE) is reduced by 42%~72% compared to MPC, highlighting the potential of DRL in lane-keeping tasks.

To improve the accuracy and stability of trajectory tracking control for intelligent vehicle, the paper presents an extension coordinated control approach that simultaneously improves tracking performance and maintains yaw stability in the presence of model uncertainty. In the upper layer, the real-time vehicle driving state is mapped to the extension set based on extension theory, where the control mode and weight distribution are determined according to the correlation function. The lower layer considers modeling errors and utilizes a trajectory tracking controller that integrates an Extended State Observer (ESO) with Model Predictive Control (MPC) to compute the front wheel steering angle control input. The Active Disturbance Rejection Control (ADRC) theory is employed to formulate a Direct Yaw Moment Controller (DYC), calculating the corresponding wheel braking torque according to the desired yaw rate. Simulation results validate that the improved tracking performance and stability using the proposed approach.

To address the issue of motion sickness among passengers in autonomous driving scenarios, which significantly compromises the travel experience, this study proposes an olfactory and auditory mitigation method based on EEG analysis. Using mint gas and alpha brainwave music as stimuli, an autonomous driving environment is constructed via a driving simulation platform. EEG signals are collected from 30 participants, and three motion sickness mitigation protocols (auditory, olfactory, and combined auditory-olfactory) are designed. Combining temporal and frequency domain features (mean, variance, skewness, kurtosis, power spectral density) from EEG signals with classification algorithms, an EEG-based motion sickness recognition model is developed. This model is then used to evaluate the efficacy of the three approaches in alleviating motion sickness. Results indicate: The auditory & olfactory protocol demonstrated the highest efficacy, reducing Relief Index (RI) by an average of 58%. The olfactory-only protocol showed moderate efficacy, reducing RI by an average of 35%. The auditory-only protocol exhibited the lowest efficacy, reducing RI by an average of 21%.