基於視覺分析之鐵路平交道違規事件偵測

Abstract

近年來由於用路人的不當行為造成鐵路平交道事故頻傳。本論文提出一個基於電腦視覺的平交道路口汽車、機車及行人行為分析系統。此系統首先利用高斯混合模型進行前景偵測，而後對於偵測的結果進行ㄧ些後處理以消除雜訊以及減少前景破碎。接著我們衡量兩張影像前景的外接矩形之距離關係以進行追蹤，並且透過幾條判定違規準則的訂定，來檢視每條代表物件行為的軌跡是否異常(違規)。經由幾個真實場景實驗後的統計，本論文提出的鐵路平交道違規分析系統之正確率可達90%以上。

Recently the number of collision accidents at level crossings has increased due to road and railway user's improper/illegal behavior. In this thesis, we present a computer vision-based system to analyze the motion of vehicles, bikers and pedestrians. To analyze the uncharacteristic motions, this system first performs robust object detection, by using the Gaussian mixture models (GMM) to construct background model and segment foreground regions. Additionally, we provide some post-processing method to reduce noise and foreground fragments. Bounding box of each foreground region is then tracked using the distance between each bounding box of the object in the current frame and each object that was tracked in the previous frames. Finally, we establish some rules to check whether the behavior associated with each track is a violation. Experimental evaluations of the proposed approach show that an accuracy rate of more than 90 % can be achieved with the proposed approach.