Recognition Algorithm Research On The Moving Target In Dynamic Scenes

Recognition technology on moving target in dynamic scenes as an important branch of machine vision has widely using field. The pedestrian detection is one of the fields. With the technology development of image and pattern recognition, pedestrian detection has come to our life from theory research. It could be widely used in smart surveillance, driver assistant system, smart android and so on. At the present time, pedestrian detection mostly concludes two classes. The one is knowledge-based approach, the other is machine learning-based approach. Especially, the integrated learning is a hot topic in the machine learning area.A knowledge-based algorithm was proposed firstly in this thesis. It selects body symmetry and head shoulder profile as pedestrian's feature because these features are unique and have certain robustness. And then, detect pedestrian in the background of campus road by searching body symmetry and matching head shoulder profile.For the poor generalization of knowledge-based pedestrian detection algorithm, a machine learning-based algorithm was proposed. Aiming at the drawback of Dalal's HOG feature that block is too small, only to describe detail feature, this thesis proposes a kind of variable dimension HOG feature that not only present detail feature, but also present total and local feature. And aiming at too large dimension of this feature, employ Fisher criterion for selecting high discrimination feature to represent pedestrian. Finally, employ RBF-SVM to classify. The result of experiment shows that this improved algorithm can accelerate Dalal's algorithm from 1s to 397ms in comparable classification accuracy.Aiming at unbalance classes and sensitive cots in Dalal's algorithm improved, cascade structure and risk sensitive SVM are employed respectively. For training any cascade series, employ method based Fisher criterion prejudgment and method based Adaboost integrated learning respectively. The experiment on INRIA pedestrian dataset shows that detection rate can reach to 91% and 94% respectively while the false detection rate both is 1/1000. For the image of 320*240, our method can process average 8～10 frame per second.