چکیده
کلید واژه ها
1. مقدمه
1.1. انگیزه
1.2. مشارکت
1.3. طرح کلی
2. مطالعات مرتبط
2.1. راه حل های کلاسیک مبتنی بر یادگیری ماشین
2.2. راه حل های پیشرفته مبتنی بر یادگیری عمیق
3. DCNN-TFO: شبکه عصبی کانولوشن عمیق برای نقاط پرت جریان ترافیک
3.1. تجزیه
3.2. شبکه عصبی پیچشی
3.3. مدل فیوژن
3.4. شبه کد
4. ارزیابی عملکرد
5. بحث و گفتگو
6. چشم اندازهای آینده
7. نتیجه گیری
منابع
Abstract
Keywords
1. Introduction
1.1. Motivation
1.2. Contribution
1.3. Outline
2. Related work
2.1. Classical machine learning based solutions
2.2. Advanced deep learning based solutions
3. DCNN-TFO: Deep convolution neural network for traffic flow outliers
3.1. Decomposition
3.2. Convolution neural network
3.3. Fusion model
3.4. Pseudo-code
4. Performance evaluation
5. Discussions
6. Future perspectives
7. Conclusion
Declaration of Competing Interest
References
چکیده
این مقاله یک معماری یادگیری عمیق جدید را برای شناسایی نقاط پرت در زمینه سیستمهای حمل و نقل هوشمند ارائه میکند. استفاده از یک شبکه عصبی کانولوشن با یک استراتژی تجزیه کارآمد برای یافتن رفتار غیرعادی دادههای جریان ترافیک شهری مورد بررسی قرار گرفته است. مجموعه دادههای جریان ترافیک شهری به خوشههای مشابهی تجزیه میشود که هر کدام حاوی دادههای همگن است. شبکه عصبی کانولوشنال برای هر خوشه داده استفاده می شود. به این ترتیب، مدلهای مختلفی آموزش داده میشوند که هر کدام از دادههای بسیار همبسته آموخته شدهاند. در نهایت از یک استراتژی ادغام برای ترکیب نتایج مدل های به دست آمده استفاده می شود. برای تایید عملکرد چارچوب پیشنهادی، آزمایشهای فشرده بر روی دادههای جریان ترافیک شهری انجام شد. نتایج نشان میدهد که سیستم ما در چندین معیار دقت از رقبا بهتر عمل می کند.
توجه! این متن ترجمه ماشینی بوده و توسط مترجمین ای ترجمه، ترجمه نشده است.
Abstract
This paper presents a novel deep learning architecture for identifying outliers in the context of intelligent transportation systems. The use of a convolutional neural network with an efficient decomposition strategy is explored to find the anomalous behavior of urban traffic flow data. The urban traffic flow data set is decomposed into similar clusters, each containing homogeneous data. The convolutional neural network is used for each data cluster. In this way, different models are trained, each learned from highly correlated data. A merging strategy is finally used to fuse the results of the obtained models. To validate the performance of the proposed framework, intensive experiments were conducted on urban traffic flow data. The results show that our system outperforms the competition on several accuracy criteria.
Introduction
Urban traffic flow data have recently piqued the curiosity of researchers. [1–3], in particular, numerous deep learning and computer vision systems [4–6] have been implemented to analyze and understand urban traffic flow data in the context of intelligent transportation, and smart city based applications. Urban traffic flow data consists of observations such as the number and speed of cars or other vehicles at specific locations, as determined by installed sensors. These numbers represent the flow of traffic, which is related to the capacity of roads and the demand on the transportation system. Urban planners are interested in the effects of various factors on traffic flow that result in unexpected patterns called outliers. In addition, we hope to learn from the behavior of independent participants (bicyclists, cars, trucks, and public transit) under different conditions (weather, events, road maintenance) to help urban planners and managers make decisions about roadway design, regulatory systems (e.g., traffic signals), and public transit routes, as well as temporary invasive building placement decisions.
Conclusion
In this paper, we explored decomposition and deep learning to accurately retrieve the abnormal behavior of urban traffic flow data. The data is first divided into clusters, each of which contains similar data. This makes the training process of the convolutional neural network simpler and more oriented to homogeneous behaviors. As a result of this combination, several models are trained, each representing the data of the corresponding cluster. A fusion model is developed to combine the results of the trained models. Extensive testing was performed to improve the validation procedure of the proposed framework. The results obtained using urban traffic data show that the proposed approach outperforms state-of-the-art outlier identification algorithms using many metrics.