Research Of Pedestrian Detection Algorithm Based On Feature Fusion And Movement Information

With the rapid development of cities and urbanization, security issues aroused more attention. Especially in the fields of smart security, transportation, public safety, the safety of pedestrians is more important. Thus pedestrian detection has also been a widespread concerned research. As a common problem in computer vision, pedestrian detection has made great progress, but there are some difficulties: Pedestrians’ distinguishing characteristic with other objects is its non-rigid object properties, there are different pedestrian standing posture, clothing style, color and the complexity of some of the occlusion condition; Monitoring angle, changes in the light conditions, and other uncertainty in background are exit. Therefor the accurate the pedestrian detection faces many challenges.In this paper, widely used detection methods are introduced at home and abroad. The Statistic-based classification algorithms are researched, according to the general flow of pedestrian detection, feature extraction, classification algorithms are discussed. Including pedestrians popular features feature are described, also an analysis of comparison of the popular pedestrian classification algorithm was made. On this basis, this paper presents Haar-like feature-based head shoulders model to describe the characteristics of pedestrians better, give a fast calculation method for HOG feature to reduce the amount of computation, and design a HOG Classifier with SVM algorithms and a Haar-like Classifier with Ada Boost learning and classification. Eventually a kind of feature-based cascade pedestrian detection algorithm was built which has two steps: HOG Classifier is used to get possible pedestrian windows, Haar-like Classifier confirm information of head shoulders in the the result windows.In the stage of pedestrian detection on the video, several moving object detection algorithms are compared in the advantages and disadvantages. Background subtraction is chosen as aided algorithm in this paper. Introducing it will shrink the detection windows which not only reduce calculation amount, but give a precise positioning moving objects to lower false positive rate.Finally, in this paper, all experiments are performed by a common international standard data set on which feature classifiers are trained. Movement information is extracted byCodebook background model with a good operating efficiency and effect. Meanwhile, the use of Open CV with GPU acceleration support is provided in parallel computing HOG feature calculation. Experiments demonstrate that the algorithm proposed have a certain promotion in terms of both accuracy and efficiency. Certainly, there are many aspects to be refined and improved, especially in some scenarios such as scene changes greatly and pedestrian occlusions.