Pedestrian Detection In Still Images

Pedestrian detection is a hot topic in the field of machine vision in recent years. During to various appearance, different background and wide range of poses, it is still a challenging task to detect human effectively. The prevalent approach is mainly based on the machine learning method which extracts the representative features from a great many training samples and then trains a relevant classifier. The method is also adopted in this paper.Two primary aspects exists in the the pedestrian detection. One is the feature selected to represent humans, the other is the classifier training method.The powerful HOG(Histogram of Oriented Gradients)feature is introduced in this paper, which was put forward by Dalal et al in 2005. Considered that HOG feature utilizes human shape information, the paper take advantage of image texture information to represent human by LBP feature presented by Ojala T. et al. SVM algorithm is employed to train the respective classifiers. Experiment results show that the LBP feature we used have achieved equivalent performance as classical HOG feature did. Inspired by this point, LBP and HOG feature is combined to propose an augmented feature vector to better represent human.In the choice of classifier training methods, we use the modified cascaded AdaBoost algorithm for our task. As AdaBoost is kind of feature selected method, it could avoid the possible existed conflict between HOG and LBP feature. What is more, the simple thresholded classifier is fast computation.When detecting human in the static image, multi-results might be generated on one single object. Result fusion method was taken to get final detection result.The paper implemented the training and testing part using VC++6.0 and OpenCV. The detection result in the paper shows that our method has some precision advantage, for it could largely tackle the varied human poses and complex background problem.