標題: 基於Haar特徵和HoG特徵的行人偵測
Pedestrian Detection Based on Haar and HoG features
作者: 陳奕竹
Chen, Yi-Chu
林進燈
Lin, Chin-Teng
電控工程研究所
關鍵字: 行人偵測;Haar;HoG;pedestrian detection;gpu
公開日期: 2011
摘要: 行人位置偵測是行人影像監測系統的第一個步驟,此部分結果的正確性對於接下來的行人追蹤、行人姿勢辨別有高度的影響,此外系統是否需要前置準備時間,也影響到實際應用的可能性。本論文提出一個行人位置偵測演算法,可以在不事先建立背景模型的前提下,在環境中選取行人的位置,並且對其可能前進方向做預測,由於不需要限制背景環境,此初步的行人位置偵測,可使用非固定的攝影機來做偵測,不論室內或室外,都可使用以此演算法發展出的即時行人偵測系統。我們將演算法區分為兩部分:第一個部分為行人位置偵測,第二個部分為行人前進方向預測。在行人位置偵測部分,我們使用一個雙層的架構,第一層使用Haar特徵訓練出來的連結分類器,快速地挑選出畫面中有可能為行人的位置,第二層再使用HoG特徵訓練出來的行人線性支持向量機,用更精細的方法,再次確認是否為行人,並將提取出來的HoG特徵,做為下一階段進行行人前進方向預測的資料。這種修改過的雙層架構,結合了連結分類器運算時間快、可容忍的樣本間差異較大等優點,對於短時間內掃瞄出畫面中姿勢各異的行人很有幫助,且其可支援緊密掃描,對於人群密度較高,有遮蔽現象發生時,也能達到高偵測率,而第一層架構上的錯判問題,又可依靠第二層精細提取的HoG特徵來做刪減。下一個部分的行人前進方向預測,也使用第二層提取出的行人HoG特徵來做判斷,我們將行人姿勢分別為正向、背向、往左和往右四個方向,由於使用線性支持向量機來做方向的判別,我們可以在很短的時間內,估測出行人可能的前進方向,此資訊可供作為更進一步的追蹤系統的資訊,提供給使用者更有用的資訊。
此演算法有以下優點:(1)由於使用大量影像訓練而成,影像含蓋不相同的背景和光線條件,和一般以外形為基礎的方法中,需要針對輪廓建立人形模型的缺點。(2)不需要使用連續的畫面,只需要一張影像就能初步進行行人辨識處理。(3)對於行人外型變化容忍度高,身形、姿勢、衣著和手持物都不易影響偵測結果。(4)演算法速度快,可應用於即時偵測系統或是移動的攝影機中。(5)實驗結果裡,最好的行人偵測系統的辨識率達到90.31%。若實現於即時的行人偵測系統上,偵測率達到87.43%。
In this thesis, we propose a new pedestrian detection algorithm without a pre-built background model, and use it to develop a real-time pedestrian detection system. The pedestrian detection algorithm can be functionally partitioned into two parts: pedestrian detection and pedestrian orientation estimation. In pedestrian detection, we use a two-stage structure to detect pedestrians in real-time. At the first stage, we use a cascade classifiers trained with Haar features to quickly select regions of interest in the frame. At the second stage, we further confirm each region of interest with a linear support vector machine trained with HoG (Histogram of oriented gradients) features. In the pedestrian orientation estimation part, the program will use the preserved HoG features to suggest the moving direction of each pedestrian inside an incompletely controlled environment. The second stage is implemented on graphic processing unit (GPU) using CUDA techniques. CUDA further increase the processing speed of the system.
This method could be used as a preprocessing stage of a tracking system or an on-board vision system for vehicles. And may help facility surveillance cameras located in building, entrances and on road with automatic detection capability.
We combined two kinds of features trained by different style classifiers to maintain the system with short processing time and high detection rate. In pedestrian recognition, we obtain different scale of pedestrian examples from online datasets. And manually categorize the samples in different orientations. We exploit the relation of two kinds of features and classifiers. Then, experiments are detailed set to extract the proper parameters for different situation.
The pedestrian detection method has the following advantages: (1) Learning from masses of example images could expand the inter objects flexibility and adapt different environment. (2)Pedestrian recognition and orientation estimation could be done in a single image. (3)The system can detect pedestrians in variety of postures, shapes and walking directions. (4) The processing speed is fast. This algorithm can be realized in real time. (5) From experiments, accuracy rate of recognition could achieve up to 90.31%. We implement this algorithm in a real-time system. The system can detect pedestrians in the scene in real time with 87.43% detection rate.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT079812547
http://hdl.handle.net/11536/46904
Appears in Collections:Thesis