| 摘要: | 監控複雜多元的交通系統，對於安全維護、犯罪預防、資訊紀錄相當重要。但對於現今發展快速的電腦視覺，影像監測不能單只是紀錄或人力觀察，更要應用在空間對象計數偵測、交通流量、物件追蹤及危險預防的判斷。公共空間能展現高時間高密度的活動聚集、短時間人群停留的特性，降低行人穿越的時間成本、增加危險發生時的人群疏散及潛在災害的危險預防等，公共空間的行人偵測技術應用就相當重要。因此本研究使用最基礎及有效偵測行人流量Aggregate Channel Features (ACF)、People Detector (PD)兩種方法，作為行人影像偵測計數方法的比較。
早期Navneet Dalal和Bill Triggs所發表應用於行人檢測的描述方向梯度直方圖(Histogram of Oriented Gradient, HOG)，到後期延衍生出如:聚合通道特徵、人臉偵測、紅外線影像偵測邊緣檢測、光流檢測、特徵檢測、圖像分割檢測、局部二值模式、特徵提取、立體視頻深度估計等。本研究針對大量的人流資料偵測對ACF與PD程式編碼做部分調整。再藉由台北捷運地下層探討評估ACF、PD檢測的有效性。
探討辨識技術ACF、PD方法的有效性，收集台北車站人流調查資料，並強化程式碼做多筆數檢測的速度與輸出。提出可能影響因子，建立計算因子數據的Matlab程式碼。依觀察者檢證資料檢討兩種方法對行人影像辨識的有效性。由相關係數值與回歸在人數、誤差分階分析可得知，ACF檢測相關係數分析較為穩定、高。明度及光源從迴歸檢查與兩方法的關聯性，PD方法均有關係。但ACF對明度也有關係外，光源則沒有。本次研究檢核方法提升其有效性，建構影像辨識技術之人流調查系統雛形，可提高公共空間安全度之維護。
Monitoring complex and diverse traffic systems is important for safety maintenance, crime prevention, and information logging. However, for the rapid development of computer vision, image monitoring can not only be used to store images or human observation, but also to detect spatial object detection, traffic flow, object tracking and risk prevention. Public space can show characteristics of long-term and high-density activity gathering, short-term crowd staying, reduce the time cost of pedestrian crossing, increase the evacuation speed of people when danger occurs, risk prevention in potential disasters, etc. Therefore, the application of pedestrian detection technology in public spaces is quite important. This study used two methods of Aggregate Channel Features(ACF) and People Detector(PD). And these are the most basic and effective methods for detecting pedestrian traffic as a comparison of pedestrian image detection and counting methods.
In the early Navneet Dalal and Bill Triggs published a description of the Histogram of Oriented Gradient (HOG) applied to pedestrian detection. Later extensions such as: aggregation channel features, face detection, infrared image detection edge detection, optical flow detection, feature detection, image segmentation detection, local binary patterns, feature extraction, etc. Partial adjustment of ACF and PD code encoding to perform a big data of pedestrian flow data detection. Then explore the effectiveness of ACF and PD detection through the underground layer of the Taipei MRT.
Discuss the effectiveness of the identification techniques ACF and PD. Collect the survey data of the Taipei station's pedestrian flow, strengthen the speed and output of the code data to do multiple data detection. Propose the possible influence factors and establish the Matlab code for calculating the factor data. Relevant analysis and regression are assessed through relevant influencing factors and various factors. From the correlation analysis of the numerical segmentation of the error and pedestrian number. It can be known that the correlation analysis of the ACF detection is relatively stable and high. Use regression to check the brightness and correlation between the light source and the PD method. They are related to the PD method. ACF also has a correlation with brightness, but the light does not. The research method of this research has improved its effectiveness, and the prototype of the image recognition technology human flow measurement system can improve the maintenance of public space security. |
