| In recent years, pedestrian detection as an important branch of object detection, which has get more and more attention and research in the field of computer vision, has become a hot topic. Pedestrian detection could be widely used in intelligent transportation, control systems, motion analysis, games, entertainment and other fields. It can not only be applied directly to the actual production and life in the past, but also to lay the foundation for human tracking and identification technologies such as computers, it has a high research value.This article focuses on the study of describing characteristics operators of the feature image as well as the model of target detection, this paper takes the form of semantics, with different combinations of forms and rules to describe some of the models. For example, from a simpler just type model to complex deformable component model, and gradually to enrich pedestrian semantic model. Then to improve the capability of pedestrian detection through training the model after establishing a semantic model, and finally to validate the corresponding model in the designated PASCAL VOC Challenge dataset.Since there are various limitations in HOG features, we need to introduce another characterization operator--HSC feature, which uses sparse learning ways to acquire the target image features, and become more robust and discriminating. The paper used is a “weak” label which can not indicate the deformable components,in order to train the weak label data set, this paper introduces the weak label structure SVM(WL-SSVM) method, which compared with the traditional method, the performance has been greatly improved.In the actual study of pedestrian detection, the detection of pedestrian target is extremely strict, it requires not only high accuracy, but also faster. This article taken cascade testing mode to choose the cascade thresholds and cutting threshold, which compared with the non-cascade method, the time-consuming is reduced to 1/4. |
PDF Full Text Request