CAJ | 학술논문

为提高行人检测的准确率，提出基于分割集成的方法用于静态图片中的行人检测。先将每个训练样本均匀分割成若干区域，提取特征后利用AdaBoost算法对每个区域建立一个局部分类器，这些局部分类器加权组成一个全局分类器。采用不同的分割方法重复上述过程，得到多个全局分类器。为进一步提高检测效果，得到更好的平均性能，对每种分割方法分别使用方向梯度直方图、多尺度方向梯度直方图特征建立2个全局分类器。当检测新的窗口时，集成上述全局分类器，通过加权投票的方式决定最终的检测结果。在INRIA公共测试集上的实验表明，文中方法有效提高检测效果。
위제고행인검측적준학솔，제출기우분할집성적방법용우정태도편중적행인검측。선장매개훈련양본균균분할성약간구역，제취특정후이용AdaBoost산법대매개구역건립일개국부분류기，저사국부분류기가권조성일개전국분류기。채용불동적분할방법중복상술과정，득도다개전국분류기。위진일보제고검측효과，득도경호적평균성능，대매충분할방법분별사용방향제도직방도、다척도방향제도직방도특정건립2개전국분류기。당검측신적창구시，집성상술전국분류기，통과가권투표적방식결정최종적검측결과。재INRIA공공측시집상적실험표명，문중방법유효제고검측효과。
To improve the accuracy of pedestrain detection, an ensemble approach for pedestrian detection in still images is proposed. Firstly, a partition ensemble method is used to evenly split the entire training window to get small regions, and features of small regions are extracted. Then, the AdaBoost classifiers are trained on different regions to get part classifiers. A global classifier is formed by weighted summing of these part classifiers. More global classifiers are obtained by using different partitioning methods to repeat the process. To improve detection results and achieve better performance, two global classifiers are built by using histograms of oriented gradient, and multi-level version of HOG descriptor features respectively for each partitioning method. The classifier ensemble is used to detect new images and the weighted voting method is used to decide the final results. Experimental results show that the proposed method achieves better performance than the whole window detector on INRIA dataset.