透過深度學習基於影像之交通特性提取

Image-based Traffic Characteristics Extraction through Deep Learning

王仁駿(Jen-Chun Wang)

指導教授 ： 陳柏華

國立臺灣大學/工學院/土木工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900418

摘要

本研究提出一個基於深度學習，利用影像以提取交通特性的架構，目的是要為交通領域提供新的視野，並期許能藉由資料分析，幫助解決交通問題。過去使用的影像方法大多著重於特徵提取並搭配分類器以辨識車輛，而深度學習方法近年興起，且在過去較少被應用於交通領域。又近年來基於深度學習的物體偵測方法進步快速，且學習能力強、網路架構能彈性調整、偵測效果佳、同一偵測器可適用於不同場景等優勢，因此利用深度學習，期望能夠有效地應用於交通問題上。藉由近年出現的基於深度學習的物體偵測方法–You Only Look Once (YOLO)，針對其具有偵測速度快 (接近即時)、偵測準確度高、背景錯誤率低、能偵測多類別物體等特性，故選用此方法作為本研究之偵測模式。本文提出之研究架構，先使用 COCO及MIO-TCD資料集訓練客製化YOLO二代模型，之後搭配卡曼濾波器及匈牙利演算法，偵測及追蹤汽車、機車、公車、自行車等物體類別，並取得分別對應的軌跡，亦可估算各個類別之數量，於單一張顯示卡NVIDIA GTX 1080上可達大約 38 FPS，汽車分類之精確率可達91.46%以上，機車可達89.51%以上。本文提出之研究架構應用於各種場景皆能保持一定的偵測精確率，又偵測速度快，且不受相機鏡頭、天氣狀況影響，可應用於許多交通問題。

關鍵字

電腦視覺 ； 深度學習 ； 物體偵測 ； 物體追蹤 ； 交通特性提取

並列摘要

We proposed an image based structure to extract traffic characteristics using deep learning. The purpose is to provide new perspectives for the traffic engineering field and hope to facilitate the analysis related to traffic engineering problems. In the past, image based methods mostly focused on feature extraction combined with classifiers to detect and recognize vehicles. In recent years, deep learning methods have risen while there are not many applications in the field of traffic engineering, and deep learning based object detection methods have progressed rapidly. It has the advantages of strong learning ability, flexibility of network architecture, good detection effect, and robustness to different scenes. Therefore, it is possible that the image-based deep learning method can be effectively applied to traffic engineering problems. You Only Look Once (YOLO), a deep-learning-based object detection method, has the advantages of fast detection speed (almost real-time), high detection accuracy, low background error rate, and the ability to detect multiple categories of objects. Due to these characteristics, it was chosen as the detector of this study. We used the COCO and MIO-TCD datasets to train the customized YOLO v2 model. Together with the Kalman filter and the Hungarian algorithm, this work can precisely detect and track objects such as cars, motorbikes/scooters, buses, and bicycles, and obtain the corresponding trajectory and counting of each category of objects, with the frame rate up to about 38 FPS on a single graphics card NVIDIA GTX 1080. The classification precision can achieve over 89.51% for motorbikes, 91.46% for cars. This model can be applied to various scenes to maintain a certain detection accuracy rate, and also has a fast detection speed, not affected by the camera lens and weather conditions, and can be applied to many traffic problems.

並列關鍵字

Computer Vision ； Deep Learning ； Object Detection ； Object Tracking ； Traffic Characteristics Extraction

參考文獻

[1] A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft, “Simple online and realtime tracking,” Proc. - Int. Conf. Image Process. ICIP, vol. 2016–Augus, pp. 3464–3468, 2016.

[2] Z. Sun, G. Bebis, and R. Miller, “On-road vehicle detection: A review,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 5, pp. 694–711, 2006.

[3] C. Tzomakas, W. VonSeelen, I. Neuroinformatik, and R. Bochum, “Vehicle Detection in Tra c Scenes Using Shadows Vehicle Detection in Tra c Scenes Using Shadows 2 Detecting Shadows,” no. August, 1998.

[4] A. Bensrhair, M. Bertozzi, A. Broggi, P. Miche, S. Mousset, and G. Toulminet, “A cooperative approach to vision-based vehicle detection,” IEEE Intell. Transp. Syst., pp. 209–214, 2001.

[5] Z. Sun, R. Miller, G. Bebis, and D. DiMeo, “A real-time precrash vehicle detection system,” Proc. IEEE Work. Appl. Comput. Vis., vol. 2002–Janua, pp. 171–176, 2002.