使用保守驗證之適應性線上Boosting偵測器

Abstract

在基於影像的即時車輛偵測中，AdaBoost演算法已經被廣泛的使用。然而，AdaBoost演算法訓練出的偵測器，為了偵測不同環境的車輛，需要大量的影像樣本來增加車輛的偵測率及降低誤判率。同時也表示，它的適性越廣，整個偵測器規模就會越大，導致偵測速度下降。 本研究提出應用on-line boosting解決AdaBoost演算法為了適應不同環境所產生的偵測器過載現象。主要使用AdaBoost演算法維持較好的車輛偵測效果及以On-line Boosting演算法降低偵測誤判率。我們提出的方法有以下優點: (1)降低車輛偵測器中弱分類器的總數，使偵測時處理速度變快。(2)降低偵測系統收斂的時間。

Adaboost algorithm is widely used in video-based real-time vehicle detection recently. However, to deal with the detection in different scene, AdaBoost algorithm requires huge data to train a detector with high detection rate and low false alarm rate. Accounting various environments, the adaboost detector is getting complex with high computational load.. Hence, this paper applied on-line boosting to resolve the overloading of AdaBoost detector for local bias of environments. We combine AdaBoost detection model with On-line Boosting model. We use AdaBoost model to detect most vehicles in high detection rate, and on-line boosting model to deal with false alarm arising in different environments. The proposed method has the following advantages: (1) Reducing number of features in the detector to deal with different environments, the processing speed become faster. (2) The proposed system can converge in a short time.