多攝影機監測路口多目標人員移動之研究

A Study on Monitoring the Movement of Multi-Personnel at an Intersection with Multi-Camera

廖育賢(LIAO, YU HSIEN)

指導教授 ： 廖俊鑑

朝陽科技大學/資訊學院/資訊與通訊系/碩士(2020年)

摘要

隨著道路汽車數量快速的增長，交通意外更是頻繁的發生。根據世界衛生組織報告中指出，道路上最常發生意外的是路上的行人、腳踏車及摩托車騎士，而死亡意外最嚴重的往往也發生在上述族群。故本研究旨在使用多攝影機對路口進行監測，並追蹤道路上的移動目標，藉以達到讓道路使用者瞭解到路口的使用狀況。
本文提出使用多攝影機對單一道路口，對多個目標人員的監測方法。使用多攝影機是為提供更多不同角度的影像來源，藉以減少在行駛及因其他物體所造成的遮蔽死角。多目標偵測的方法使用You Only Look Once的方法來進行，可偵測出道路上的最常見的物體，結合Simple Online and Realtime Tracking的方法可對多個偵測目標進行追蹤，紀錄目標移動軌跡，接著使用透視轉換對多個影像來源中的目標進行座標的轉換，並使用K-means進行多目標在不同來源影像上，結合重複偵測到的目標，並顯示在一張俯瞰圖中，達成對單一道路中的多個目標進行監測。
藉由本研究的監測方法，可以使實際使用者瞭解到道路的使用情形。實驗中證明本論文使用的方法可以在實際道路中，達成七成的偵測準確率，並對多個目標進行持續的追蹤。

關鍵字

深度學習 ； 目標偵測 ； 目標追蹤 ； 道路監測

並列摘要

With the rapid increase more vehicles on the road, traffic accidents are more frequent. According to the World Health Organization report, the most frequent accidents on the road are pedestrians, bicycles and motorcyclists, and the most serious accidents often occur in the above-mentioned groups. Therefore, this study aims to use multi-camera to monitor intersections and track moving objects on the road, so that road users can understand the use of intersections and make judgments about possible accidents.
This paper proposes the use of multi-camera to monitor a single intersection and multiple objects personnel. Using multi-camera to provide more sources of images at different angles. As to reduce the blind spots caused by driving and other objects. The objects detection method uses You Only Look Once to detect the most common object on the road. Combined with the method of Simple Online and Realtime Tracking, it can track detected objects and record the moving track of the objects. Using perspective transformation to transform the coordinates of objects in multiple image sources. Using K-means to combine objects on different source images and repeat the combination of perspectives. Displayed in a top view to achieve monitoring of objects on a single road.
With the monitoring method of this study, actual users can get a better understanding of road usage. The experiment proves that the method used in this paper can achieve a 70% detection accuracy rate on actual roads, and continuously track multiple targets.

並列關鍵字

Deep learning ； Object detection ； Object tracking ； Road monitoring

參考文獻

[1]交通部公路總局統計查詢網。(2020, July 15) [online] Available:https://stat.thb.gov.tw/hb01/webMain.aspx?sys=100&funid=11100

[2]內政部警政署交通事故統計。(2020, July 15) [online] Available:http://stat.motc.gov.tw/mocdb/stmain.jsp?sys=100&funid=b3303

[3]Global status report on road safety 2018. (2020, July 15) [online] Available: https://www.who.int/violence_injury_prevention/road_safety_s tatus/2018/en/

[4]S. Singh, “Critical Reasons for Crashes Investigated in the National Motor Vehicle Crash Causation Survey,” Traffic Safety Factors Crash Stats, National Center for Statistics and Analysis, Washington D.C., USA, February, 2015.

[5]杜絕馬路三寶、車禍，通訊系統「V2X」要讓車子開口說話、溝通。(2020, July 15) [online] Available: https://www.bnext.com.tw/article /51397/itri-v2x