Traffic light detection with YOLO models
Application of advanced artificial intelligence methods to improve driving, increase safety and efficiency of road traffic. Performance evaluation of improved "You Only Look Once" image set models developed for traffic light detection and recognition.
| Рубрика | Коммуникации, связь, цифровые приборы и радиоэлектроника |
| Вид | статья |
| Язык | английский |
| Дата добавления | 05.09.2024 |
| Размер файла | 20,7 K |
Отправить свою хорошую работу в базу знаний просто. Используйте форму, расположенную ниже
Студенты, аспиранты, молодые ученые, использующие базу знаний в своей учебе и работе, будут вам очень благодарны.
Размещено на http://www.Allbest.Ru/
Lviv Polytechnic National University, Ukraine
Department of Artificial Intelligence
Traffic light detection with YOLO models
Yu. Zanevych, Bachelor's degree student
Summary
The accurate detection and recognition of traffic lights are paramount in the realm of autonomous driving systems and intelligent traffic management. This study leverages the comprehensive cinTA_v2 Image Dataset on Robot flow, specifically designed for traffic light detection, to evaluate the performance of advanced You Only Look Once (YOLO) models, including YOLOv7l, YOLOv8n, YOLOv8s, and YOLOv8m. Through meticulous training and evaluation, we systematically analyses the models' ability to accurately detect and classify traffic light states (green, red, and yellow) under a variety of challenging conditions. Our findings reveal significant improvements in precision, recall, and mean Average Precision (mAP) across the models, with YOLOv8m demonstrating superior overall performance, especially in terms of mAP50-95, reflecting its enhanced capability in detecting small and partially obscured traffic lights. The study not only showcases the effectiveness of YOLO models in a critical application within the autonomous driving domain but also highlights the potential for further advancements in traffic light detection technologies. By discussing the challenges, limitations, and future directions, this work contributes to the ongoing efforts to improve road safety and efficiency through the application of cutting-edge artificial intelligence techniques.
Keywords: Autonomous Driving, Traffic Light Detection, YOLO Models, YOLOv7l, YOLOv8n, YOLOv8s, YOLOv8m, Object Detection, Computer Vision, Deep Learning, Image Recognition, cinTA_v2 Dataset, Robotflow, Precision and Recall, Mean Average Precision (mAP).
Introduction
In the burgeoning field of autonomous driving and intelligent transportation systems, the ability to accurately detect and recognize traffic lights is not just a technical requirement but a cornerstone of safe and efficient navigation [6,7]. The march towards fully autonomous vehicles has intensified the need for sophisticated computer vision techniques capable of interpreting complex environments with precision [9]. Among the myriad of technologies developed for this purpose, the You Only Look Once (YOLO) models stand out as a formidable tool for real-time object detection, celebrated for their speed and accuracy [4, 5, 8, 10]. However, applying YOLO models specifically for traffic light detection presents a unique set of challenges. These include but are not limited to varying lighting conditions, the different states of traffic lights, and occlusions, all of which can significantly impact the performance of detection algorithms.
Accurate recognition of traffic lights is a pivotal task within the broader context of autonomous navigation, demanding high precision and recall to safeguard the interaction between autonomous systems and human drivers. The stakes are high, as errors in detection can lead to traffic violations or accidents, undermining the safety and trust in autonomous technologies [6, 11]. Despite remarkable advancements in autonomous vehicle technologies, achieving robust traffic light detection in diverse and dynamic environments remains a daunting challenge. It is a challenge that not only tests the limits of current technologies but also pushes the envelope for innovation in intelligent transportation systems.
This study aims to address these challenges head-on by employing the cinTA_v2 Image Dataset on Robot flow [1], a meticulously curated dataset designed to enhance the performance of traffic light detection algorithms. Through a detailed evaluation of several YOLO models, including YOLOv7l, YOLOv8n, YOLOv8s, and YOLOv8m, this research end eavors to benchmark their capabilities in accurately detecting and classifying traffic lights under a variety of conditions. By leveraging this comprehensive dataset, our research not only benchmarks the current capabilities of these models but also sheds light on areas ripe for improvement, setting the stage for the development of more reliable autonomous driving technologies.
Moreover, the findings of this study have far-reaching implications for the development of intelligent transportation systems at large, emphasizing the critical importance of continuous advancements in object detection methodologies. As we navigate through the complexities of real-world scenarios, the insights garnered from this research underscore the indispensable role of deep learning in enhancing the safety and efficiency of autonomous driving systems. In doing so, we contribute to the body of knowledge on traffic light detection, paving the way for future innovations that will continue to shape the landscape of autonomous driving and intelligent transportation systems.
Methodology
Dataset Overview. The foundation of our study is the cinTA_v2 Image Dataset on Robotflow, a robust collection designed to advance traffic light detection technology. This dataset encompasses a total of 2,397 images, each meticulously annotated to include labels for three traffic light states: green, red, and yellow. These images were captured in a variety of environmental conditions, including different times of day, weather scenarios, and urban settings, to mimic the complexities encountered in real-world driving situations [1].
To ensure a thorough evaluation of the YOLO models, the dataset was strategically divided into training, validation, and testing sets. The training set comprises 87% of the total images, amounting to 2,097 images. This extensive collection is used to train the models, allowing them to learn and adapt to the diverse scenarios presented by the dataset. The validation set consists of 8% of the images, totaling 200 images, which is utilized to fine-tune model parameters and prevent overfitting. Finally, the testing set makes up 4% of the dataset, with 100 images, providing a basis for unbiased evaluation of the models' performance in detecting and classifying traffic light signals [1].
This structured dataset split supports a rigorous training and evaluation process, facilitating an accurate assessment of the models' ability to recognize traffic lights under various challenging conditions. It underscores our commitment to leveraging detailed, real-world data to enhance the development of autonomous driving technologies.
Model Overview. In our study, we scrutinize the performance of four distinct YOLO (You Only Look Once) models: YOLOv7l, YOLOv8n, YOLOv8s, and YOLOv8m. These models represent the forefront of object detection technology, each offering unique advantages in terms of accuracy, speed, and computational efficiency.
• YOLOv7l is considered for its balanced performance, serving as a robust baseline for traffic light detection. Its architecture is designed to offer a compromise between detection speed and precision, making it suitable for real-time applications
• YOLOv8n, the first of the YOLOv8 series evaluated in our study, introduces optimizations in model architecture that aim to reduce computational load while maintaining high detection accuracy [3]. It is particularly designed for environments where computational resources are limited.
• YOLOv8s is a step up in complexity and performance from YOLOv8n, aiming to strike a balance between the computational efficiency of YOLOv8n and the higher accuracy of larger models [3]. It is tailored for scenarios that demand both highspeed processing and precision.
• YOLOv8m represents the most advanced model in our study, boasting the highest accuracy among the evaluated models. It incorporates more complex architectural features and larger parameter sets, making it well-suited for applications where detection precision is paramount [3].
Each model's design and capabilities are leveraged to address the specific challenges presented by traffic light detection, such as small object sizes, occlusions, and variable lighting conditions.
Training Process. The training of each YOLO model was meticulously planned and executed to ensure optimal performance on the cinTA_v2 Image Dataset [1]. We employed a uniform training protocol across all models to maintain consistency in the evaluation process.
Data Augmentation: To enhance the models' ability to generalize across various conditions, we implemented a series of data augmentation techniques, including random scaling, cropping, flipping, and brightness adjustments. These augmentations simulate diverse lighting and environmental conditions, preparing the models for real-world traffic light detection tasks.
Hyperparameter Tuning: The selection of hyperparameters was carried out through a series of preliminary experiments designed to identify the optimal configurations for learning rate, batch size, and the number of epochs. This process was guided by the objective to maximize the models' precision and recall while ensuring efficient training times.
Model Selection Criteria: The criteria for model selection were based on a combination of performance metrics, including precision, recall, and mean Average Precision (mAP). Special emphasis was placed on mAP50-95, a rigorous metric that evaluates detection accuracy across a range of Intersection Over Union (IoU) thresholds. This comprehensive assessment ensures that the selected models demonstrate robust performance in accurately detecting and classifying traffic lights under varied conditions.
Through this systematic training process, we aimed to not only assess the current capabilities of YOLO models in traffic light detection but also to identify areas for further improvement and optimization.
artificial intelligence traffic light recognition road safety
Results
Our comprehensive evaluation of YOLOv7l, YOLOv8n, YOLOv8s, and YOLOv8m models on the cinTA_v2 Image Dataset has yielded significant insights into their performance in traffic light detection tasks. The results are discussed below, focusing on precision, recall, and mean Average Precision (mAP) metrics across all classes (green, red, and yellow).
Overall Performance. Across the board, all models demonstrated high levels of precision and recall, indicating their effectiveness in traffic light detection:
• YOLOv7l showed remarkable precision (0.96) and recall (0.949), with an mAP of 0.979, setting a strong baseline for comparison.
• YOLOv8n offered a slightly lower precision (0.947) but improved recall (0.962), achieving an mAP50 of 0.986 and an mAP50-95 of 0.622.
• YOLOv8s further enhanced precision (0.973) and recall (0.966), matching the mAP50 of YOLOv8n and surpassing its mAP50-95 with a score of 0.634.
• YOLOv8m emerged as the top performer, with the highest precision (0.989), recall (0.984), and an impressive mAP50 of 0.992 and mAP50-95 of 0.639.
Class-wise Performance. The models' abilities to detect specific traffic light states revealed nuanced strengths and weaknesses:
• For green traffic lights, YOLOv8n achieved the highest precision (0.999), while YOLOv8m led in recall (0.964) and closely matched precision (0.987), showcasing their effectiveness in green light detection.
• Red traffic lights saw YOLOv8m at the forefront with both the highest precision (0.991) and recall (0.986), indicating its superior ability to detect red signals.
• Yellow traffic lights were most accurately detected by YOLOv8s and YOLOv8m, both achieving perfect recall (1.0). YOLOv8m had the highest precision (0.989), slightly edging out YOLOv8s.
Comparative Analysis. The comparison highlights YOLOv8m's overall superiority in detecting traffic lights, demonstrating advancements in YOLO model architectures and training methodologies. While YOLOv7l provided a solid baseline, the progression to YOLOv8 variants, especially YOLOv8m, indicates significant improvements in both precision and recall, which are crucial for the real-world application of these models in autonomous driving systems.
The class-wise analysis underscores the models' varying capabilities in distinguishing between different traffic light states, with YOLOv8m consistently showing exceptional performance. This suggests that the architectural and algorithmic enhancements in YOLOv8m are particularly effective in addressing the challenges posed by traffic light detection, such as small object sizes and varying illumination conditions.
Discussion
The evaluation of YOLOv7l, YOLOv8n, YOLOv8s, and YOLOv8m models on the cinTA_v2 Image Dataset has provided insightful results into the capabilities and limitations of state-of-the-art object detection models in traffic light detection tasks.
Our findings underscore the substantial progress made in applying deep learning to real-world challenges, yet they also highlight areas where further research and development are necessary.
Model Efficacy. The YOLOv8m model demonstrated superior performance across all metrics, indicating its enhanced capability to detect and classify traffic lights accurately. This model's success can be attributed to its complex architecture and larger parameter set, which, despite requiring greater computational resources, offers significant improvements in detection precision. The advancements in YOLOv8 series models, as evidenced by their performance on the cinTA_v2 dataset, reflect the ongoing refinement of neural network architectures for object detection tasks.
However, the nuanced differences in performance between models, particularly in class-wise detection, point to an interesting observation: there is no one-size-fits-all model for traffic light detection. The varying conditions under which traffic lights must be detected--ranging from bright daylight to dim night scenes, and from clear to occluded views--necessitate a model that can adapt to a wide range of scenarios. The higher precision and recall rates of YOLOv8m for detecting specific colors of traffic lights suggest that further optimizations in model architecture could yield even better results, especially in challenging conditions.
Challenges and Limitations. A critical challenge encountered in this study relates to the inherent limitations of the dataset and model training processes. While the cinTA_v2 Image Dataset offers a diverse set of images, the dynamic and unpredictable nature of real-world driving conditions means that even more varied datasets could enhance model robustness. Additionally, the trade-off between model complexity and computational efficiency remains a significant consideration. Models like YOLOv8m, although highly accurate, may not be practical for real-time applications on hardware with limited processing power.
Future Directions. This study illuminates several pathways for future research. Firstly, the exploration of hybrid models or ensemble techniques could potentially leverage the strengths of different architectures to improve detection accuracy. Secondly, integrating other forms of sensory data, such as LIDAR or radar, with image-based detections could offer a more holistic approach to traffic light detection, especially in adverse weather conditions or when direct line-of-sight is obstructed. Lastly, advancements in training methodologies, such as few-shot learning or transfer learning, could enable models to achieve high levels of accuracy with less data, addressing the challenges of dataset diversity and availability.
Conclusion
This study embarked on an evaluation of the latest YOLO models - YOLOv7l, YOLOv8n, YOLOv8s, and YOLOv8m - utilizing the comprehensive cinTA_v2 Image Dataset to gauge their effectiveness in the detection and classification of traffic lights. The results underscored the superior performance of the YOLOv8m model across various metrics, highlighting its enhanced capability to accurately detect and classify traffic lights in diverse environmental conditions. Such findings are not only testament to the advancements in deep learning architectures and their applicability in real-world scenarios but also spotlight the critical role of high-quality, diverse datasets in training and evaluating object detection models.
Our research reveals that while significant strides have been made in the field of traffic light detection, challenges remain, particularly in balancing model complexity with computational efficiency and ensuring robustness across all traffic light states and conditions. The nuanced performance differences among the evaluated models emphasize the need for continued innovation in model architecture and training methodologies.
Looking forward, the study illuminates several avenues for further research. The exploration of more complex model architectures, hybrid models, or ensemble techniques could offer new paths to enhance accuracy and reliability. Moreover, integrating multimodal data sources and adopting advanced training techniques may address current limitations and unlock new potentials in traffic light detection technology, ultimately contributing to the safety and efficiency of autonomous driving systems.
In conclusion, the advancements in YOLO models and their application to traffic light detection highlight the ongoing evolution of autonomous driving technologies. As we continue to refine these models and methodologies, the vision of fully autonomous vehicles navigating complex urban environments safely and efficiently becomes increasingly attainable. The findings from this study not only contribute to the academic and practical knowledge base but also pave the way for future innovations in the field of intelligent transportation systems.
References
1. Pradana, W. (2022). cinTA_v2 Dataset. Roboflow Universe. Roboflow. Retrieved March 16, 2024,
2. Wang, C.-Y., Bochkovskiy, A., & Liao, H.-Y.M. (2022). YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. arXiv.
3. Reis, D., Kupec, J., Hong, J., & Daoudi, A. (2023). Real-time flying object detection with YOLOv8. arXiv.
4. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779-788.
5. Bochkovskiy, A., Wang, C.-Y., & Liao, H.-Y. M. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv.
6. Saini, S., Nikhil, S., Konda, K. R., Bharadwaj, H. S., & Ganeshan, N. (2017). An efficient vision-based traffic light detection and state recognition for autonomous vehicles. 2017 IEEE Intelligent Vehicles Symposium (IV).
7. Levinson, J., Askeland, J., Dolson, J., & Thrun, S. (2011). Traffic light mapping, localization, and state detection for autonomous vehicles. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA).
8. Zhao, Z.-Q., Zheng, P., Xu, S.-t., & Wu, X. (2019). Object detection with deep learning: A review. arXiv.
9. Behrendt, K., & Novak, L. (2017). A Deep Learning Approach to Traffic Lights: Detection, Tracking, and Classification. IEEE International Conference on Robotics and Automation (ICRA), 1370-1377.
10. Omar, W., Lee, I., Lee, G., & Park, K. M. (2020). Detection and localization of traffic lights using YOLOv3 and stereo vision. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B2-2020, 1247-1252.
11. Wang, X., Han, J., Xiang, H., Wang, B., Wang, G., Shi, H., Chen, L., & Wang, Q. (2023). A lightweight traffic lights detection and recognition method for mobile platform. Drones, 7(5), 293.
Размещено на Allbest.Ru
...Подобные документы
Рассмотрение понятия, основных задач и структуры биллинговых систем. Определение назначения, функциональных возможностей и преимуществ использования автоматизированных систем расчетов MS ISA Server, UТМ компании NETUP, StarGazer и Traffic Inspector.
контрольная работа [32,1 K], добавлен 14.09.2010Вимоги до транспортної мережі NGN. Порівняльний аналіз технологій транспортних мереж: принцип комутації, встановлення з'єднання, підтримка технології QoS, можливості масштабування мережі. Поняття про Traffic Engineering. Оптимізація характеристик мереж.
дипломная работа [4,6 M], добавлен 22.09.2011Технологии построения сетей передачи данных. Обоснование программных и аппаратных средств системы передачи информации. Эргономическая экспертиза программного обеспечения Traffic Inspector. Разработка кабельной системы волоконно-оптических линий связи.
дипломная работа [2,2 M], добавлен 24.02.2013Signal is a carrier of new information for the observer. Concept and classification detector signals, their variety and functional features. The detection abilities of different detector’s types, methodology and milestones of their determination.
контрольная работа [1,1 M], добавлен 27.04.2014История создания сенсорных экранов и понятие их технологии в наши дни. Суть технологий IntelliTouch и Acoustic Pulse Recognition (APR). Взаимодействие экрана и оператора ввода, принцип действия устройства. Преимущества и недостатки сенсорных экранов.
реферат [1,9 M], добавлен 10.05.2013Ознайомлення з технологією 4G та її комунікативна характеристика, можливості та особливості, вимоги IMT-Advanced. Карта покриття України та перспективи. Зростання пристроїв з підтримкою LTE. Порівняльна характеристика 4G і 3G, використовуване обладнання.
реферат [1,4 M], добавлен 06.11.2016Обзор технологий LTE, действующих в мире. Параметры стандарта LTE Advanced (Rel.10). Основные положения радио доступа. Расширения, добавленные в стандарт. Разделение контрольной информации и данных. Расчёт зоны покрытия базовых станций сети LTE Rel.8.
курсовая работа [2,4 M], добавлен 25.02.2015История использования подводных лодок ВМФ США. Описание боевых информационно-управляющих систем (БИУС) на их борту как комплекса электронно-вычислительной аппаратуры для управления и эффективного использования боевых и технических возможностей оружия.
презентация [896,8 K], добавлен 23.12.2013Основні можливості пакету Image Processing. Дослідження методів перетворення цифрових зображень в середовищі Matlab. Відновлення розмитого зображення за допомогою команди deconvblind, його геометричні перетворення. Зашумлення зображення функцією motion.
курсовая работа [2,4 M], добавлен 05.02.2015Initial data for the term paper performance. Order of carrying out calculations. Analyze uncompensated system. Synthesize the real PD-compensator ( ) which would guarantee desired phase margin at gain crossover frequency . Analyze compensated system.
курсовая работа [658,7 K], добавлен 20.08.2012Изучение метода цифровой трассерной визуализации Stereo Particle Image Velocimetry (Stereo PIV), предназначенного для измерения трехкомпонентных полей скорости в выбранном сечении потока. Анализ основных конфигураций для стереоскопических измерений.
контрольная работа [592,4 K], добавлен 01.02.2017Использование для построения модели сети сухопутной подвижной связи технологии IMT Advanced, которая относится к четвертому поколению мобильной связи. Расчет частотно-территориального планирования, построение модели блока системы подвижной связи.
курсовая работа [871,7 K], добавлен 16.02.2013Concept and functional features of piezoelectric sensors, the scope of its application. Designing with piezoelectric sensors. Piezo-vibration sensor Parallax 605–00004 and Bosch 608–00112: overview, technical characteristic, accessories, installations.
контрольная работа [1,1 M], добавлен 27.05.2013Create a source of light in Earth orbit. Energy source for the artificial sun. Development of light during nucleosynthesis. Using fusion reactors. Application lamp in the center of a parabolic mirror. Application of solar panels and nuclear reactors.
презентация [2,7 M], добавлен 26.05.2014Technical methods of supporting. Analysis of airplane accidents. Growth in air traffic. Drop in aircraft accident rates. Causes of accidents. Dispatcher action scripts for emergency situations. Practical implementation of the interface training program.
курсовая работа [334,7 K], добавлен 19.04.2016Definition and classification of marketing communications, their variety and comparative characteristics. Models of formation of enterprise marketing, evaluation of their efficiency, structure and components. Factors influencing consumer behavior.
презентация [2,7 M], добавлен 25.11.2015Impact of globalization on the way organizations conduct their businesses overseas, in the light of increased outsourcing. The strategies adopted by General Electric. Offshore Outsourcing Business Models. Factors for affect the success of the outsourcing.
реферат [32,3 K], добавлен 13.10.2011Review of development of cloud computing. Service models of cloud computing. Deployment models of cloud computing. Technology of virtualization. Algorithm of "Cloudy". Safety and labor protection. Justification of the cost-effectiveness of the project.
дипломная работа [2,3 M], добавлен 13.05.2015Basic assumptions and some facts. Algorithm for automatic recognition of verbal and nominal word groups. Lists of markers used by Algorithm No 1. Text sample processed by the algorithm. Examples of hand checking of the performance of the algorithm.
курсовая работа [22,8 K], добавлен 13.01.2010Метафора и метонимия и их механизмы. Анализ языковых и речевых значений прилагательных "heavy", "light" и "weak" с целью вывода их лексических прототипов. Разновидности устной речи. Факторы, влияющие на успешность развития навыков говорения у школьников.
дипломная работа [951,1 K], добавлен 21.01.2017
