Reasearch On Vedio Images-Based Pedestrian Abnormal Behaviors

Studies on pedestrians' abnormal behaviors are both a hot topic in the field of computer vision and also a crucial technique problem to realize intelligent monitoring in the application branch of video monitoring.Although a large number of achievements on pedestrians' abnormal behavior detection have been reported for the past decades,the related research development is slow in that it is difficult to extract pedestrians' behavior characteristics in complex scenarios.Therefore,in order to investigate the problem of pedestrians' abnormal behavior presented in video monitoring images,this academic dissertation aims at studying background modeling algorithms on foreground target extraction and developing feature vector models which present pedestrians' local and global feature information,while it also probes into pedestrians' abnormal detection approaches.The main works and acquired achievements are summarized as follows:A.An improved ViBe target detection approach,which associates with a novel background update strategy,is proposed to handle the problems of pedestrian foreground extraction and shadow elimination in video images.In the design of the algorithm,an improved background modeling approach is designed to solve the problem of pedestrian foreground extraction in given static environments,after some statistical characteristics on adjacent frame pixel intensity and time distribution are combined with the basic ViBe background modeling approach;meanwhile,an improved shadow elimination algorithm is also designed to deal with those shadows in video images,based on the HSV color space and the luminance relation between foreground targets and shadow regions.Comparatively numerical experiments show that the target detection approach,with satisfactory shadow elimination and low computational complexity,can effectively extract foreground targets in video images.B.For the problem of design on pedestrians' behavior feature vector modeling,such several feature models as pedestrians' profile feature,motion velocity,profile change curve and local frequency detection,etc.,are designed to describe the performances of president's gesture andlocal and global movement.These models,together with the Hu moment features,derive a feature model to reflect pedestrian's behavior status.After that,a template-based abnormal behavior recognition algorithm is designed to realize the recognition of abnormal behavior of pedestrians in video images,relying upon a behavioral template library.Finally,the comparative experiments validate that one such abnormal behavior recognition approach is superior to the compared algorithms.C.After the conventional BP neural network is analyzed and summarized with the aspects of structural design,algorithm principle and learning rate design,an improved BP neural network is acquired to deal with the problem of pedestrians' abnormal behavior recognition after an adaptive correction rule on learning rate is designed.Numerical experiments illustrate that,whereas the improved neural network has not distinct superiority over the conventional BP neural network with the aspect of behavior recognition,it can stably carry out pedestrians' abnormal detection.