Zikai Yao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zikai Yao has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Qiang Liu
School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China
4 shared publications - Jinglong Zhu
School of Intelligent Systems Engineering, Sun Yat-sen University, Shenzhen, China
3 shared publications Qiang Liu
Sun Yat-sen University
3 shared publicationsQing Li
3 shared publications- Zhangzhen Zhao
School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China
3 shared publications Bo Li
Research and Development Center, Guangzhou Automobile Group Company Ltd., Guangzhou, China
2 shared publications
Latest Publications
- 31 July 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (30) , 306107
A unified domain adaptation framework for traffic light recognition in adverse scenes
Traffic light recognition under adverse scenes is a fundamental yet challenging task for intelligent driving systems, where domain shift caused by variations in weather, illumination, and geographic conditions often leads to significant performance degradation. To address this issue, this paper proposes a unified domain-adaptive traffic light recognition framework, termed DA-TLR, to improve cross-domain generalization and robustness. Unlike existing methods that treat feature alignment in a decoupled manner, DA-TLR formulates domain adaptation as a unified modeling problem across both image-level and instance-level discrepancies. Specifically, DA-TLR is built upon the YOLOv5-l architecture and integrates a teacher–student module, an image-level adaptation module, and a domain adaptation loss for joint optimization. The teacher–student module employs a knowledge distillation strategy to generate high-confidence pseudo labels, facilitating instance-level feature alignment. The image-level adaptation module combines bidirectional pseudo-image transformation and dynamic masking to reduce appearance discrepancies and enhance robustness against domain-specific variations. Extensive experiments on multiple domain adaptation scenarios, including foggy, rainy, nighttime, and cross-regional conditions, demonstrate that DA-TLR consistently outperforms existing methods. In particular, the proposed method achieves mAP 50 scores of 89.57% and 79.03% in cross-fog and cross-rain tasks, respectively, approaching the upper bound of target-domain supervised training. Further analysis indicates that jointly modeling multi-level domain discrepancies is more effective than treating them independently. Overall, DA-TLR provides an effective and practical solution for robust traffic light recognition under adverse scenes, and offers a promising direction for domain-adaptive perception in real-world intelligent driving systems.
- 31 July 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (30) , 306107
A unified domain adaptation framework for traffic light recognition in adverse scenes
Traffic light recognition under adverse scenes is a fundamental yet challenging task for intelligent driving systems, where domain shift caused by variations in weather, illumination, and geographic conditions often leads to significant performance degradation. To address this issue, this paper proposes a unified domain-adaptive traffic light recognition framework, termed DA-TLR, to improve cross-domain generalization and robustness. Unlike existing methods that treat feature alignment in a decoupled manner, DA-TLR formulates domain adaptation as a unified modeling problem across both image-level and instance-level discrepancies. Specifically, DA-TLR is built upon the YOLOv5-l architecture and integrates a teacher–student module, an image-level adaptation module, and a domain adaptation loss for joint optimization. The teacher–student module employs a knowledge distillation strategy to generate high-confidence pseudo labels, facilitating instance-level feature alignment. The image-level adaptation module combines bidirectional pseudo-image transformation and dynamic masking to reduce appearance discrepancies and enhance robustness against domain-specific variations. Extensive experiments on multiple domain adaptation scenarios, including foggy, rainy, nighttime, and cross-regional conditions, demonstrate that DA-TLR consistently outperforms existing methods. In particular, the proposed method achieves mAP 50 scores of 89.57% and 79.03% in cross-fog and cross-rain tasks, respectively, approaching the upper bound of target-domain supervised training. Further analysis indicates that jointly modeling multi-level domain discrepancies is more effective than treating them independently. Overall, DA-TLR provides an effective and practical solution for robust traffic light recognition under adverse scenes, and offers a promising direction for domain-adaptive perception in real-world intelligent driving systems.
- 26 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (6) , 9448-9461
Vehicle Trajectory Prediction Based on Dynamic Spatiotemporal Density-Adaptive Fusion in Autonomous Driving
In complex driving environments, accurately predicting vehicle trajectories is crucial for the safe operation of intelligent vehicles. This paper introduces a Dynamic Spatiotemporal Density-Adaptive Fusion Network (DSDAF) designed to address the challenges of multi-modal trajectory prediction in dynamic traffic scenarios. To capture the spatiotemporal dependencies between vehicles, we utilize the multi-head attention mechanism to model their dynamic interaction. We further propose a dynamic density-aware adaptive mechanism that employs a transformer-based local vehicle density prediction network to quantify future traffic states and dynamically integrate them with historical trajectory information, thereby enabling adaptive prediction under real-time traffic conditions. Additionally, we propose an interpretable decoder architecture based on Kolmogorov-Arnold Networks (KAN), which decomposes driving behaviors into lateral and longitudinal components. By leveraging learnable activation functions and dynamic attention mechanisms, our model generates more accurate intention representations and multimodal trajectories. Extensive experiments conducted on two real-world datasets demonstrate that the proposed method outperforms state-of-the-art algorithms, particularly in long-term predictions, while ablation studies further verify the effectiveness of each component.
- 26 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Vehicular Technology
- Vol. 75 (6) , 9448-9461
Vehicle Trajectory Prediction Based on Dynamic Spatiotemporal Density-Adaptive Fusion in Autonomous Driving
In complex driving environments, accurately predicting vehicle trajectories is crucial for the safe operation of intelligent vehicles. This paper introduces a Dynamic Spatiotemporal Density-Adaptive Fusion Network (DSDAF) designed to address the challenges of multi-modal trajectory prediction in dynamic traffic scenarios. To capture the spatiotemporal dependencies between vehicles, we utilize the multi-head attention mechanism to model their dynamic interaction. We further propose a dynamic density-aware adaptive mechanism that employs a transformer-based local vehicle density prediction network to quantify future traffic states and dynamically integrate them with historical trajectory information, thereby enabling adaptive prediction under real-time traffic conditions. Additionally, we propose an interpretable decoder architecture based on Kolmogorov-Arnold Networks (KAN), which decomposes driving behaviors into lateral and longitudinal components. By leveraging learnable activation functions and dynamic attention mechanisms, our model generates more accurate intention representations and multimodal trajectories. Extensive experiments conducted on two real-world datasets demonstrate that the proposed method outperforms state-of-the-art algorithms, particularly in long-term predictions, while ablation studies further verify the effectiveness of each component.
- 25 June 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 294
FIGS-SLAM: Gaussian splatting SLAM with dynamic frequency control and influence-based pruning
- 25 June 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 294
FIGS-SLAM: Gaussian splatting SLAM with dynamic frequency control and influence-based pruning
- 2 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 55 (6) , 4410-4422
Night-Time Traffic Light Recognition Based on Enhancement-Guided Object Detection
Traffic light recognition is crucial for autonomous driving. While significant progress has been made in favorable conditions, recognition performance in night-time scenes remains a challenge. One straightforward approach is to apply enhancement methods that improve degraded images prior to object detection. However, since most enhancement methods are tailored for human perception; they may not consistently improve recognition accuracy for machine learning techniques. To address this, we propose an enhancement-guided framework for night-time traffic light recognition, called EG-TLR. EG-TLR consists of a residual denoising module (RDM) and a mixed attention traffic light detection module (MATLDM). The RDM reduces noise in degraded night-time images while preserving essential traffic light features by extracting sparsity information and performing context aggregation. The MATLDM improves feature extraction and recognition performance in complex night-time scenes by incorporating a shadow detection layer (SDL) and a mixed attention module (MAM). Moreover, to address the lack of a dedicated night-time traffic light dataset, we construct the Night-TL dataset utilizing publicly available images. Extensive experiments on Night-TL and LISA datasets demonstrate that EG-TLR achieves an AP50 of 79.83% and an AP50:95 of 36.62%, with an inference speed of 5.9 ms and 16.9 GFLOPs, outperforming other state-of-the-art methods. Furthermore, ablation studies and visualization results validate the effectiveness of our proposed method. The Night-TL dataset can be downloaded from: https://github.com/feiqinaqian/Night-TL-dataset.
- 2 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 55 (6) , 4410-4422
Night-Time Traffic Light Recognition Based on Enhancement-Guided Object Detection
Traffic light recognition is crucial for autonomous driving. While significant progress has been made in favorable conditions, recognition performance in night-time scenes remains a challenge. One straightforward approach is to apply enhancement methods that improve degraded images prior to object detection. However, since most enhancement methods are tailored for human perception; they may not consistently improve recognition accuracy for machine learning techniques. To address this, we propose an enhancement-guided framework for night-time traffic light recognition, called EG-TLR. EG-TLR consists of a residual denoising module (RDM) and a mixed attention traffic light detection module (MATLDM). The RDM reduces noise in degraded night-time images while preserving essential traffic light features by extracting sparsity information and performing context aggregation. The MATLDM improves feature extraction and recognition performance in complex night-time scenes by incorporating a shadow detection layer (SDL) and a mixed attention module (MAM). Moreover, to address the lack of a dedicated night-time traffic light dataset, we construct the Night-TL dataset utilizing publicly available images. Extensive experiments on Night-TL and LISA datasets demonstrate that EG-TLR achieves an AP50 of 79.83% and an AP50:95 of 36.62%, with an inference speed of 5.9 ms and 16.9 GFLOPs, outperforming other state-of-the-art methods. Furthermore, ablation studies and visualization results validate the effectiveness of our proposed method. The Night-TL dataset can be downloaded from: https://github.com/feiqinaqian/Night-TL-dataset.
- 25 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 26 (7) , 10295-10307
MADFM: Multi-Layer Adaptive Driver Fatigue Monitoring Model Based on Steering Wheel Signals
Driver fatigue poses a serious threat to road safety. Although most vehicles are equipped with driver fatigue monitoring models, the performance of existing models is not satisfactory due to the complexity and unclear reliability of fatigue features. To address these issues, this study proposed a multi-layer adaptive driver fatigue monitoring (MADFM) model, based on the steering wheel signals obtained from 25 participants in a driving simulator. The steering wheel signals were sequentially processed through an adaptive window partitioning algorithm based on sample entropy, feature extraction, and an adaptive outlier point removal algorithm based on local outlier factor. Finally, a multi-scale adaptive feature selection (AFS) algorithm based on global-local correlation was applied to obtain the optimal feature set Do. AFS algorithm was then validated using a random forest model and compared with existing feature selection algorithms. The results showed that after adopting the AFS algorithm, the model’s fatigue monitoring accuracy reached 93.71%, which was 10.21% higher than the algorithm without using feature selection. Compared to four common feature selection algorithms, AFS algorithm demonstrated superior performance in terms of monitoring accuracy, precision, recall and interpretability. The interpretability analysis results revealed that steering wheel angle and lateral acceleration features were the most sensitive features affecting the accuracy of the fatigue monitoring model. This model could improve the monitoring performance of driver fatigue, prompting safer driving behaviors, and reducing traffic accidents risk.
- 25 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 26 (7) , 10295-10307
MADFM: Multi-Layer Adaptive Driver Fatigue Monitoring Model Based on Steering Wheel Signals
Driver fatigue poses a serious threat to road safety. Although most vehicles are equipped with driver fatigue monitoring models, the performance of existing models is not satisfactory due to the complexity and unclear reliability of fatigue features. To address these issues, this study proposed a multi-layer adaptive driver fatigue monitoring (MADFM) model, based on the steering wheel signals obtained from 25 participants in a driving simulator. The steering wheel signals were sequentially processed through an adaptive window partitioning algorithm based on sample entropy, feature extraction, and an adaptive outlier point removal algorithm based on local outlier factor. Finally, a multi-scale adaptive feature selection (AFS) algorithm based on global-local correlation was applied to obtain the optimal feature set Do. AFS algorithm was then validated using a random forest model and compared with existing feature selection algorithms. The results showed that after adopting the AFS algorithm, the model’s fatigue monitoring accuracy reached 93.71%, which was 10.21% higher than the algorithm without using feature selection. Compared to four common feature selection algorithms, AFS algorithm demonstrated superior performance in terms of monitoring accuracy, precision, recall and interpretability. The interpretability analysis results revealed that steering wheel angle and lateral acceleration features were the most sensitive features affecting the accuracy of the fatigue monitoring model. This model could improve the monitoring performance of driver fatigue, prompting safer driving behaviors, and reducing traffic accidents risk.
- 29 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 25 (10) , 13887-13899
A Coarse-to-Fine Deep Learning Based Framework for Traffic Light Recognition
Accurate recognition of traffic light plays a crucial role in ensuring road traffic safety. Due to the wide variety of traffic light classes and the inherent conflict between location determination and intra-class classification of traffic lights, existing deep learning-based methods cannot obtain satisfactory results, making them still unreliable as traffic light recognizers. To address this issue, we propose an effective Coarse-to-fine framework for traffic light recognition, which is consisted of an enhanced super class detector and a fine-grained classifier. The main contributions lie in three aspects. Firstly, we determine the optimal number of super class by adopting a clustering algorithm. It can alleviate the effect caused by similar classes of traffic light. Secondly, we enhance the super class detector by introducing an additional prediction layer, which enable utilizing the shadow features and improving the detection performance of traffic light super class. Lastly, we propose a fine-grained classifier based on depth-wise separable convolution and expansion layer for super class candidate classification. It can achieve superior recognition performance of different traffic lights especially for similar classes. To evaluate the effectiveness of our method, we conduct experiments on WPI (Worcester Polytechnic Institute) traffic light dataset. With an input size of 416 ${\times}$ 416 pixels, the proposed method can achieve a precision of 99.04% and a recall of 97.40% with a speed of 9.9ms. Compared with other state-of-the-art methods, the results demonstrate the significant competitiveness of our method and its high effectiveness for traffic light recognition.
- 29 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 25 (10) , 13887-13899
A Coarse-to-Fine Deep Learning Based Framework for Traffic Light Recognition
Accurate recognition of traffic light plays a crucial role in ensuring road traffic safety. Due to the wide variety of traffic light classes and the inherent conflict between location determination and intra-class classification of traffic lights, existing deep learning-based methods cannot obtain satisfactory results, making them still unreliable as traffic light recognizers. To address this issue, we propose an effective Coarse-to-fine framework for traffic light recognition, which is consisted of an enhanced super class detector and a fine-grained classifier. The main contributions lie in three aspects. Firstly, we determine the optimal number of super class by adopting a clustering algorithm. It can alleviate the effect caused by similar classes of traffic light. Secondly, we enhance the super class detector by introducing an additional prediction layer, which enable utilizing the shadow features and improving the detection performance of traffic light super class. Lastly, we propose a fine-grained classifier based on depth-wise separable convolution and expansion layer for super class candidate classification. It can achieve superior recognition performance of different traffic lights especially for similar classes. To evaluate the effectiveness of our method, we conduct experiments on WPI (Worcester Polytechnic Institute) traffic light dataset. With an input size of 416 ${\times}$ 416 pixels, the proposed method can achieve a precision of 99.04% and a recall of 97.40% with a speed of 9.9ms. Compared with other state-of-the-art methods, the results demonstrate the significant competitiveness of our method and its high effectiveness for traffic light recognition.
- 26 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 24 (9) , 9736-9750
TL-Detector: Lightweight Based Real-Time Traffic Light Detection Model for Intelligent Vehicles
With the leap-forward development of intelligent driving, traffic light detection with high accuracy and high speed is important for intelligent transportation systems to ensure safety. However, existing methods encounter difficulties in balancing the detection speed and accuracy. This paper aims to present a flexible and robust lightweight model TL (Traffic Light)-Detector for real-time detection of traffic light. The model is composed of three parts. An enhanced backbone network combined with G-module is proposed to generate abundant information and reduce computational load, and the coordinate attention mechanism is introduced to focus on location features thereby strengthening the feature extract ability. The lightweight neck promotes multi-scale feature information aggregation for feature fusion to build a lightweight feature fusion network. The lightweight detection head adopts anchor-free mechanism to eliminate the hyperparameters related to the anchor. The dataset TL2022 of traffic light is built based on real traffic sceneries. Experimental results show that TL-Detector has the best comprehensive performance. TL-Detector achieves a detection speed of 277 FPS and a precision of 73.24% with only 0.72 GFLOPs. The experiments on LaRA (La Route Automatisee) public traffic light dataset show the excellent generalization ability. It indicates that TL-Detector can effectively achieve accurate and real-time detection for traffic lights.
- 26 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 24 (9) , 9736-9750
TL-Detector: Lightweight Based Real-Time Traffic Light Detection Model for Intelligent Vehicles
With the leap-forward development of intelligent driving, traffic light detection with high accuracy and high speed is important for intelligent transportation systems to ensure safety. However, existing methods encounter difficulties in balancing the detection speed and accuracy. This paper aims to present a flexible and robust lightweight model TL (Traffic Light)-Detector for real-time detection of traffic light. The model is composed of three parts. An enhanced backbone network combined with G-module is proposed to generate abundant information and reduce computational load, and the coordinate attention mechanism is introduced to focus on location features thereby strengthening the feature extract ability. The lightweight neck promotes multi-scale feature information aggregation for feature fusion to build a lightweight feature fusion network. The lightweight detection head adopts anchor-free mechanism to eliminate the hyperparameters related to the anchor. The dataset TL2022 of traffic light is built based on real traffic sceneries. Experimental results show that TL-Detector has the best comprehensive performance. TL-Detector achieves a detection speed of 277 FPS and a precision of 73.24% with only 0.72 GFLOPs. The experiments on LaRA (La Route Automatisee) public traffic light dataset show the excellent generalization ability. It indicates that TL-Detector can effectively achieve accurate and real-time detection for traffic lights.
- Open Access
- 17 December 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 7, 183838-183846
Detection of Foreign Matter on High-Speed Train Underbody Based on Deep Learning
When a high-speed train is running, it is easy for foreign objects like rail-side plastic bags to enter bottom bogies, cables and equipment gaps, which affects the safety of driving. At present, the detection accuracy of such foreign objects is low. To solve this problem, the present study used the latest deep learning based object detection networks, such as SSD and Faster R-CNN, which combined with different feature extractors to build a detection network. Through data augmentation of a sample, the number of sizes was expanded. By the idea of transfer learning, the COCO dataset was used as the source data to transfer features to the new detection network and then retrained to build a train bottom plastic bag detection network based on deep learning. It was shown that through data augmentation, the precision of each combination significantly improved. Moreover, the results using combination of SSD and MobileNet obtained the fastest detection speed with an average time of 41 ms and a precision rate of 81.3%. The accuracy using combination of SSD+Inception V2 combination reached to 89.7%, with an average time of 53 ms.
- 17 December 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 7, 183838-183846
Open AccessDetection of Foreign Matter on High-Speed Train Underbody Based on Deep Learning
When a high-speed train is running, it is easy for foreign objects like rail-side plastic bags to enter bottom bogies, cables and equipment gaps, which affects the safety of driving. At present, the detection accuracy of such foreign objects is low. To solve this problem, the present study used the latest deep learning based object detection networks, such as SSD and Faster R-CNN, which combined with different feature extractors to build a detection network. Through data augmentation of a sample, the number of sizes was expanded. By the idea of transfer learning, the COCO dataset was used as the source data to transfer features to the new detection network and then retrained to build a train bottom plastic bag detection network based on deep learning. It was shown that through data augmentation, the precision of each combination significantly improved. Moreover, the results using combination of SSD and MobileNet obtained the fastest detection speed with an average time of 41 ms and a precision rate of 81.3%. The accuracy using combination of SSD+Inception V2 combination reached to 89.7%, with an average time of 53 ms.