基于Co-training方法的车辆鲁棒检测算法

Robust Vehicle Detection Algorithm Based on Co-training Method

针对复杂交通场景车辆检测算法自适应能力差的问题,提出了基于Co-training半监督学习方法的车辆鲁棒检测算法.首先,针对手工标记的少量样本,分别训练基于Haar-like特征的AdaBoost分类器和基于HOG(histograms of oriented gradients)特征的SVM(support vector machines)分类器,使其具有一定的识别能力;然后,基于Co-training半监督学习框架,将利用2种算法进行分类得到的新样本分别加入到对方的样本库中,增加训练样本数量,再次进行分类器的训练.由于这2类特征具有冗余性,各自检测出的正负样本包含对方漏检和误检的图像.由于样本数的增加,再次训练所得到的新分类器的鲁棒性得到了很大提高,能更加准确地检测出车辆,而且由算法对未标记样本进行分类标记,不再需要人为标记,提高了车辆检测算法的自适应能力.

To improve the adaptability of existing vehicle detection algorithms in complex traffic circumstances, a robust detection algorithm based on co-training from semi-supervised learning methods was proposed. First, according to a small number of humanly labeled samples, two classifiers were trained, which were AdaBoost based on Haar-like features and the SVM （support vector machines） based on HOG （histograms of oriented gradients） features, respectively, so that both of them had some identification ability. Second, on the basis of co-training from semi-supervised learning framework, the new samples gained from the two algorithms above were added to mutual sample sets to increase the number of training samples, and the train was repeated. Due to the redundancy these two features had, the detected positive and negative samples would contain the images which were missed out or falsely detected mutually. Because of the increasing number of samples, the robustness of the new re-training classifiers has been greatly improved so that the classifiers can detect the vehicles accurately. Besides, there will be no need to mark artificially, but to classify and mark the unlabeled samples by the algorithms. Therefore, it can highly improve the adaptability of vehicle detection algorithm.