行人偵測系統

Abstract

本論文提出一個新的行人偵測演算法，並且以此演算法發展出一套即時的行人偵測系統。這個行人偵測演算法可區分為兩部分：移動物件偵測與行人辨識。在移動物件偵測部分，我們使用改善過的時間差異法（Modified Temporal Differencing Method）在場景中切割出移動的物件，這種改善過的時間差異法結合了一般的時間差異法與偵測網路（Detection Nets）的觀念。在行人辨識部分，我們由輸入影像取得多種類型的小波樣板（Wavelet Template），再由小波樣板取出興趣點樣板（Interest Point Template），我們利用小波樣板的頻率分析及多級別（Multi-scale）特性，並且結合興趣點的區域性特徵，由興趣點樣板中萃取出統計所得的特徵點，最後，這些特徵點輸入一個訓練過的多層級倒傳遞類神經網路（Multilayer Back-propagation Neural Network），類神經網路的輸出就表示最後的結果—人或非人。此行人偵測演算法有以下幾點優點：(1)經由大量影像訓練而成，改善了一般以外形為基礎的方法中需要手繪人形模型的缺點。(2)不用求得運動型別（Motion Pattern），在行人辨識處理中只需要一張影像就可得到結果。(3)不會受到行人的衣服、髮型、性別、走路姿勢及行走方向所影響。(4)演算法速度快，可應用於即時偵測系統中。(5)經由實驗得知，此行人偵測系統的辨識率達到95%。我們將此演算法實現於即時的行人偵測系統中，行人偵測率達到89.58%。

In this thesis, we propose a new pedestrian detection algorithm, and use it to develop a real-time pedestrian detection system. The pedestrian detection algorithm can be functionally partitioned into two parts: moving object segmentation and pedestrian recognition. In moving object segmentation, we segment the moving objects in the scene by modified temporal differencing method. This method combines general temporal differencing method with detection nets. In pedestrian recognition, we obtain multi-type wavelet templates from input images. Interest points are extracted from wavelet templates. We exploit frequency analysis and multi-scale features in wavelet templates and local features in interest point templates. Then, feature points are extracted by statistical method from interest point templates. Finally, these feature points are fed into a trained multilayer back-propagation neural network. The output of the neural network implies the result – pedestrian or non-pedestrian. The pedestrian detection algorithm has the following advantages: (1) Learning from masses of example images: this avoids the drawback of general shape-based methods that need handcrafted models. (2) Not relying on motion patterns: the pedestrian recognition procedure only needs one image. (3) The system can detect pedestrians in variety of clothes, hairstyles, gender, walking poses and walking directions. (4) The processing speed in pedestrian detection is fast. This algorithm can be realized in real time. (5) From experiments, accuracy rate of recognition achieves 95%. We implement this algorithm in a real-time pedestrian detection system. The system can detect pedestrians in the scene in real time. The pedestrian detection rate is 89.58%.