運用複雜度及亮度分佈檢測之建立在Boosting基礎上的車輛偵測

Abstract

近年來，為了提高分析大量影像資料的效率與準確度，影像式的車輛偵測技術廣泛應用在智慧型運輸系統中，且相關研究以及應用也愈來愈多。目前使用最廣泛的方式是利用背景減出前景的方式來偵測目標物，然而在都市區域交通流量大且車輛走走停停的區域，使用建立背景方式的準確度會大幅下降，此外，也需要時間建立背景，而無法即時反應目前的車輛偵測結果。在本研究中，我們提出了一個可長時間並且穩定的自動化車輛偵測系統，利用AdaBoost訓練一個可適用在白天以及傍晚的車輛分類器，此分類器擁有高偵測率，但誤判率也相對比較高，且被誤判的內容也相對複雜，因此我們建立了一套可針對白天以及傍晚兩種情況去濾除誤判的系統，我們利用車體在不同光線條件下的特徵，配合計算量小的演算法來有效的濾除誤判，並同時維持整體系統運做速度，使其可以運用在即時應用上，此外，我們也透過切換機制將兩種濾除方式合而為一，來達到自動偵測的目的。最後，系統利用生存演算法將偵測結果做進一步的穩定與呈現，並且應用在車輛計數的應用上。對公開的MIT CBCL資料庫測試，我們比較了使用PCA + ICA[18]、AdaBoost + PDBNN[29]以及我們提出的方法，PCA + ICA的偵測率為95%，誤報率為0.002%，AdaBoost + PDBNN的偵測率為91.93%，誤報率為0.0031%，我們的系統可達到96.27%的偵測率，0.0015%的誤報率；也同時對我們蒐集的影片做測試，包含白天及傍晚的影片，比較對象有使用GMM建背景[34][35]和AdaBoost + PDBNN這兩種方式，使用背景方式的平均偵測率為80.7%，平均誤報率35.3%，AdaBoost + PDBNN的平均偵測率為72.5%，誤報率為7%，我們的系統的平均偵測率為98%，平均誤報率為3%。由實驗結果可以看出，我們提出的系統可以適用於現實環境中，並且可以達到即時偵測的效果。

In recent years, visual-based vehicle detection techniques have been extensively applied to Intelligent Transportation System (ITS) to improve the efficiency and precision of analyzing massive video information. The most common method is using background image to extract foreground objects. However, this method is not suitable for urban area where has heavy traffic, and vehicles will move and stop frequently. The correctness of detection will decrease dramatically. What’s more, background method cannot detect instantaneous situation of traffic because it need learning time to construct background image. In this study, we proposed a long-term and stable automatic vehicle detection system. We used AdaBoost algorithm to train a vehicle classifier which can operate both at daytime and evening. The classifier has high detection rate, but the false alarm rate is also relatively high, and the content of false alarms is complicate too. Therefore, we also developed two false alarm eliminating algorithms to reduce false alarm rate for daytime and evening respectively. We utilized the features of vehicle under different lighting conditions. Then, these features cooperate with algorithms, which need low computation power, to filter out false alarms efficiently, and maintain reasonable operating speed to let the system can be applied to real-time applications. Furthermore, we used switching algorithm to combine the two false alarm eliminating algorithm into one system to achieve automatic detection. In the end, we use survival algorithm to further stabilize and present the detection results, and applied the system to real-time vehicle-counting application. For the MIT CBCL database, we compared our system with PCA + ICA[18] and AdaBoost + PDBNN[29] these two methods. The detection rate of PCA + ICA is 95% and false alarm rate is 0.002%. The detection rate of AdaBoost + PDBNN is 91.93% and false alarm rate is 0.0031%. The detection rate of our system is 96.27% and false alarm rate is 0.0015%. For proprietary testing videos which contain both daytime and evening videos, we compared with GMM[34][35] and AdaBoost + PDBNN these two methods. The average detection rate of GMM is 80.7% and average false alarm rate is 35.3%. The average detection rate of AdaBoost + PDBNN is 72.5% and average false alarm rate is 7%. The average detection rate of our system is 98% and false alarm rate is 3%. We can tell from experimental results that our proposed system can operate in real world environment, and has real-time detection ability in the same time.