- 學位論文
應用於嵌入式駕駛輔助平台之快速行人與機車偵測追蹤系統
Fast Pedestrian and Motorcycle Detection and Tracking for Embedded Driver Assistance System
摘要
在行人與機車偵測中，先計算影像中每個像素的梯度，並且利用內插的方式將各種不同的梯度向量量化為九個固定方向的向量，將其累積以形成Histogram of oriented gradients(HOG)特徵，為了減少計算量，將其特徵建立成積分影像，以做為偵測行人與機車的特徵。當取出欲偵測區域的HOG特徵時，利用積分影像的特性對其四個角落進行加減，即可快速取得該區域之HOG特徵，然後首先使用較為快速的Cascade SVM分類器來篩選輸入的資料，通過篩選的區域將會由精確的SVM分類器來進行分類判斷是否為行人或機車。最後為了減少因外在因素使得分類器失效對系統的影響程度，使用mean shift的方式對偵測的結果做追蹤，主動去找尋偵測失敗的行人或機車。 本論文將系統移植至DSP嵌入式系統平台，使用的嵌入式平台是DM648，CPU是C64x+ 783MHz，使用C語言最佳化程式碼，許多嵌入式平台都缺乏浮點運算能力，因此做定點化將浮點數改為定點數、查表法，藉此來提高運算速度，本論文之演算法在嵌入式平台上執行1張frame的時間平均為65 ms，偵測率達到9成5，準確率也達9成以上。證明該演算法是適用於實際情境中。
並列摘要
Fast detection of pedestrians moving across the roads is a big challenge for in-vehicle embedded systems. Because the shape features of on-road pedestrians are irregular and complex, so that the detection techniques cost large computational resources. However, the in-vehicle embedded systems only have limited computational resources. The histogram of oriented gradients (HOG) is a feature descriptor to represent objects with complex shapes appearing in an image to assist computers to learn and recognize such objects and usually combined with the machine learning methods, ex. support vector machines (SVMs), boosting classifiers and decision trees. In this thesis, we propose fast object detection algorithms based on HOGs and we divide the all procedure into several modules and develop corresponding optimization approaches. The proposed techniques are evaluated and implemented on a digital signal processing (DSP) based embedded platform. Experimental results demonstrate that the proposed detection techniques can provide high computational efficiency and detection accuracy.