Research On Technologies Of Pedestrian Detection In Intelligent Patrol Robot

In recent years,along with the advance in technology social development,Intelligent video surveillance system is more and more popular.pedestrian as one of the focus of surveillance video,which is necessary to study their detection and tracking.In the unmanned driving,security,intelligent transportation and rescue and relief work field,pedestrian detection plays an indispensable role.As a significant branch of computer vision,pedestrian detection is always the focus of research.In real life,because of the complexity of background context,the variability of outdoor lighting,the non-rigid of posture and the diversity of appearance,which challenges the pedestrian detection.Subject to different application scenarios of the pedestrian detection,yet,to this day it is not a quite generic algorithm to solve the pedestrian detection in all scenarios.The title of this thesis is from the enterprise project "patrol robot",which investigates the pedestrian detection under the supervision street video,focuses on the detection precision and algorithm in real time.The main work and research contents are as follows:First of all,for the non-rigid of the pedestrian,studies various algorithms of describing features in depth,which emphatically analyze feature method of Haar,Edgelet and classification algorithm of Support Vector Machine and Adaboost.In order to solve the problem of missing pedestrians in partial occlusion and improve the accuracy of the proposed algorithm,based on the classical HOG algorithm,the HOG feature and the LBP feature are selected to be mixed,and the fused HOG-LBP feature is proposed.Through selecting INRIA and self sampling training samples,the scheme can detect the partial occlusion of pedestrians and improve the accuracy.Then,the real-time method of pedestrian detection is studied in detail.Including the use of a fast algorithm for HOG computation based on integral image,the ViBE algorithm based on background subtraction,the thesis proposes a fast pedestrian detection method based on feature selection of head and shoulder,the body's head and shoulders as target recognition of pedestrians,avoiding general feature extraction feature dimension computation a large quantity,which can quickly detect multi-scale zoom to pedestrian target.Finally,a simple pedestrian detection system is designed,and the structure of the system,software and hardware are introduced briefly.The accuracy,real-time and effectiveness of the algorithm are further verified.