Transformers have emerged as a foundational paradigm in autonomous driving, enabling high-capacity modeling of complex, multimodal, and dynamic environments. Their self-attention mechanisms, scalability, and sequence modeling capabilities support spatial–temporal reasoning and long-range dependency capture. Although increasingly adopted in core modules—such as perception, trajectory prediction, decision-making, and anomaly detection—a system-level survey of their architectural evolution and deployment challenges remains lacking. This paper presents a structured survey of Transformer models in autonomous driving, introducing a task-oriented taxonomy that spans object detection, sensor fusion, trajectory forecasting, motion planning, and intent prediction. We compare Transformer architectures with traditional deep learning models (e.g., CNNs, RNNs), highlighting advantages in global context modeling, multimodal alignment, and unified representations across heterogeneous inputs (camera, LiDAR, radar, HD maps). Beyond current applications, this work examines large-scale, end-to-end Transformer systems and their potential as foundation models for autonomous driving. We analyze design patterns involving chain-of-thought reasoning, neuro-symbolic integration, federated learning, and privacy-preserving edge deployment. Case studies from industry leaders (e.g., Tesla, Baidu, NVIDIA, Aurora) illustrate practical trade-offs and architectural adaptations. Despite progress, challenges remain in achieving real-time efficiency, robustness in open-world scenarios, interpretability in sequential decision-making, and integration with cost-sensitive sensors. We identify research gaps and propose directions in scalable Transformer design, explainable AI, and policy-aware planning. This survey aims to guide researchers and practitioners at the intersection of AI and intelligent transportation, supporting the development of interpretable, efficient, and generalizable Transformer-based autonomous driving systems.

Recommender systems (RecSys) have become critical tools for enhancing user engagement by delivering personalized content across diverse digital platforms. Recent advancements in large language models (LLMs) demonstrate significant potential for improving RecSys, primarily due to their exceptional generalization capabilities and sophisticated contextual understanding, which facilitate the generation of flexible and interpretable recommendations. However, the direct deployment of LLMs as primary recommendation policies presents notable challenges, including persistent latency issues stemming from frequent API calls and inherent model limitations such as hallucinations and biases. To address these issues, this paper proposes a novel offline reinforcement learning (RL) framework that leverages imitation learning from LLM-generated trajectories. Specifically, inverse reinforcement learning is employed to extract robust reward models from LLM demonstrations. This approach negates the need for LLM fine-tuning, thereby substantially reducing computational overhead. Simultaneously, the RL policy is guided by the cumulative rewards derived from these demonstrations, effectively transferring the semantic insights captured by the LLM. Comprehensive experiments conducted on two benchmark datasets validate the effectiveness of the proposed method, demonstrating superior performance when compared against state-of-the-art RL-based and in-context learning baselines. The code can be found at https://github.com/ArronDZhang/IL-Rec.

In human-centric applications, a common requirement is to extract information about the people present in an environment. In this paper, we introduce a static human sensing method using MIMO FMCW radar, i.e., to detect human presence and location when a person is stationary. The primary challenge for radar-based static human sensing is to separate human radar signals from environmental background. Conventional methods use signal processing techniques and handcrafted features to identify radar signals reflected from human body, which has high limitation on signal quality. We propose a deep learning model named SHSNet that automatically learn human features in radar signals from data. Our experimental results show that our method has superior accuracy in detecting human presence, identifying empty room, and localizing individuals.

Non-contact cardiac monitoring using radar technology is crucial for advancing health monitoring capabilities in smart transportation environments. This paper introduces a robust method for accurately extracting cardiac signals from radar data, specifically addressing the challenge of respiratory interference. Our approach utilizes a bank of Gabor filters, tuned to the seismocardiogram (SCG) frequency band, to effectively minimize respiratory artifacts and isolate clear heartbeat waveforms. Signal quality is further improved by selectively integrating high-fidelity cardiac data extracted from multiple locations using different filters. We validated this method with data collected from 34 participants seated up to 1 meter from the radar. Experimental results demonstrate a average median inter-beat interval (IBI) error of 20ms, with over 80% of IBI errors below 50ms.

FMCW radar-based human keypoints detection has been increasingly used in health monitoring, human-computer interaction, safety detection and other fields due to its characteristics of contact-free, privacy-preserving and less environmentdependence. Recently, most of studies have adopted deep learning method to detect human keypoints for its powerful feature extraction ability. However, its excellent performance relies on large-scale labeled datasets participating in model training, while the data labeling process is difficult and time-consuming, making it unsuitable for promotion and application. To address this issue, we propose a two-branch semi-supervised method to improve the accuracy and generalization of keypoints detection through using both labeled and unlabeled radar data. Models of two branches deal with different radar point clouds generated by same human activity, and promote the consistency and complementarity in the learning features of two branches through introducing a consistency loss function. Optimal performance in human keypoints detection can be achieved through adjusting weights between consistency loss and supervised loss. In addition, we develop a grouping processing technique of radar point clouds to obtain two different point cloud clusters with minimal feature overlap. We evaluate our model on a real-world radar dataset and compare its performance with a supervised method that only uses 20 labeled subjects. The experimental results demonstrate that the proposed method achieves a 10.7% improvement in accuracy compared to traditional supervised learning method, highlighting the significant advantages of our method in human keypoints detection and its substantial potential for broad application and promotion.

Human key-points play a vital role in smart home, elderly-care, gaming, etc. The detection methods based on millimeter wave (mmWave) radars have attracted substantial attentions and been applied in various fields because of the contactless and no privacyinvasion characteristics. In this paper, we present a machining learning-based framework, Mikey, that utilizes the commercial frequency-modulated continuous-wave (FMCW) radar point cloud to estimate the key-points. The framework consists of three steps, including point cloud adaptive merging pre-processing, key-point detection and key-point post-processing. The point cloud adaptive merging pre-processing method not only addressed the challenges stemming from the point sparsity and specular reflection but also the individual and action differences. A point cloud set is constructed adaptively from consecutive multiple frames limited by both frame and point numbers. The detection model combines the temporal and spatial features by adopting a local feature encoding block and a global self-attention block. A key-point post-processing step is added to smooth the key-point predictions and remove the jagged edge by exploring the prediction coherence in consecutive frames. We also make up for the scarcity of mmWave radar dataset for a variety of targets and actions. A dataset with 88 participants for 8 different actions is collected. We evaluate the proposed framework using additional 5 participants and obtain an average mean absolute error smaller than 6cm, confirming the effectiveness of the proposed framework.

Radar based human sensing especially vital signs(respiration/heartbeat) detection has attracted much attention. The basic principle of vital signs detection is to detect the tiny displacement caused by physiological movements, which are difficult to detect at long distance and are easily disturbed by human random body movements. To address this issue, we propose two strategies in this paper. In order not to lose the target signal, we select radar signals from multiple range bins within the neighborhood of target position as candidate signals, where the target position is obtained through examining radar signal variance. But the cost of candidate signals is to bring in noise signals that cannot be distinguished from physiological signals by using traditional signal processing methods. So a spatio-temporal dense network (ST-DenseNet) is proposed to extract physiological signals from candidate signals, which learns the most discriminative features to distinguish between physiological signals and noise signals through convolving spatial features at different temporal scales for strengthening feature fusion and using dense connections in the network for enhancing generalization capability. Based on these two strategies, our method realizes accurate vital signs detection over a large spatial range and achieves good robustness to disturbances such as body movements. Extensive experiments on wide space and multiple subjects confirms the superiority of our method. The error of respiration and heartbeat detection are reduced to 0.93bpm (beats per minute) and 3.83bpm when the sensing scope is improved to 2 meters.

Although a FMCW MIMO radar can effectively track moving people, it is a challenge for radar to detect static people. When a person is at rest, radar signals reflected from the human body are mixed up with that from stationary objects. In this paper, we present a static human localization method using FMCW MIMO radar. The method uses radar signal variations and spectral characteristics caused by vital activities to identify human locations. We test the proposed method with an off-the-shelf radar. The results show that our method can detect human presence and get human locations in a large area. Furthermore, the detected human location can be used to judge whether the target is lying on the floor, which is helpful for fall detection.

Non-contact human sensing based on radar has attracted numerous attentions and been applied in various applications, such as localization, vital sign monitoring and activity identification. This paper presents a wireless healthcare system to achieve life detection and vital sign monitoring based on frequency-modulated continuous-wave (FMCW) radar. Most of previous methods have studied to detect moving targets. The proposed life detection method has achieved stationary human target detection by utilizing the inherent characteristics of human breathing motion: spatial correlation and periodicity. Based on the phase sensitivity of FMCW radar, breathing motion is monitored within the range provided by the life detection step. Besides, different algorithms are adopted for different detection range to remove the distance and environment impacts. Experiments are carried out and have demonstrated that the proposed method can detect the stationary human targets accurately and 95% of the breathing rate estimation error is less than 3 BPM(Beat Per Minute).

The problem of characterizing voltage fluctuation is addressed as a means of the sag/swell identification in electricity networks based on the particle swarm optimization (PSO) in this paper. A new objective function of optimization algorithm to identify the voltage sags caused by short circuit fault(s) is obtained to describe the depth and types of sag/swell. Due to only a few sampling points are used in the objective function, the parameters of the ellipse could be recognized in subcycle to classify the sag. The results show that proposed method can identify not only the single stage sag but also the multistage sag within a cycle. MATLAB-based simulation results are discussed in detail to support the concept and validate the feasibility of proposed method.

This paper addresses reliable and accurate step length estimation using inertial sensors. Step length is an important parameter for the accurate position required in the location and navigation system. To tackle the challenges of drifting in accelerometer, sensitivity to user physical characteristics and walking profiles, as well as variability in environment, we have developed a calibration algorithm for reliable detection of gait parameters including step number, step frequency, maximum and minimum of acceleration magnitude. We've built a Radial-basis Function (RBF) neural network to train the model of step length that can adapt to different users. The established mathematical model can achieve the simple and efficient estimation of step length in realtime system. Extensive experiments have been conducted on 5 subjects with 3263 steps testing in total. Evaluation results showed our improved step length estimation method can achieve the recognition rate for step detection of 96% and a mean error of 0.04m for the step length estimation.

Short-range wireless positioning and inertial measurement unit (IMU) have been widely used in indoor positioning systems. However, wireless signals fluctuate seriously and the accuracy of low-cost IMU system suffers from gyroscope drift, magnetic interference and accumulative error. In this paper, we propose a fusion positioning system based on Wireless signal, Map information and Inertial sensors, which is called as WiMaIn system. The system has three key techniques, including motion recognition, particle filter and map matching. The motion state of the target is recognized from the IMU data, and then the system will choose different observations and particle transition models based on the detected motion state to get the final location estimation. Map matching is utilized to correct the heading error and constrain the weight updating of particles. Numerical experiments show that our proposed method could mitigate the accumulative error, obtain more stable performance, and achieve higher trajectory estimation accuracy than other existing methods.

Smartphones have been a great platform for location-based services (LBS) and heading estimation is a key technique. In this paper, we proposed a heading change estimation algorithm based on inertial sensors built-in smartphones for indoor positioning. Compared with previous common approaches, it gives the users a larger freedom, not requiring fixed even specific holding manners. The algorithm consists of three main parts, including holding manner transition detection, equivalent vertical angular update and heading change calculation. Holding manner transition detection aims to separate the gyroscope measurement contribution caused by transitions from that of the user's heading change. Equivalent vertical angular rate is defined to recognize the heading change interval. Due to the low-precision of inertial sensors, heading change is calculated by tracing the ideologies of complementary filter and inertial frame alignment. It can also be obtained from the equivalent vertical angular rate directly for low accuracy requirement situation. In the case of known initial heading, the proposed method is converted to a heading estimation method. Experiment results show that the proposed method can detect the holding manner transition accurately and provides relatively good heading change estimation.

Since the Quality of Experience (QoE) metric directly reflects the subjective experiences of users, the resource allocation techniques with the objective to improve QoE instead of conventional metrics are efficient in resource utilization. This kind of techniques can avoid wasting resources in providing perfect objective metric which may have little or even no effect on user experience. This paper presents a QoE-aware power allocation algorithm in multiuser Orthogonal Frequency Division Multiplexing (OFDM) systems. The power allocation problem is formulated to maximize the overall QoE subject to the total transmit power constraint. The original model is solved by means of transformation and approximation, since it is difficult to develop a closed-form solution. Simulation results show that the proposed algorithm can enhance the performance in terms of both the overall QoE and single user QoE in comparison with the classical water-filling algorithm.