Research On Methods And Application Of Fast Robust Pedestrian Detection And Tracking With Complex Scene

Pedestrian detection and tracking technique plays an important role of intelligent security video surveillance system in computer vision domain.It has profound prospect in applications of unmanned driving,video surveillance and video retrieval.Besides,the most concerning problems is targeting the pedestrian under changeable posture,mainly causing by diversity of wearing,non-rigid deformation,occlusion and other factors.This dissertation presents solution to these challenges by optimizing detection and tracking progress.Particularly,a significant progress is made with a presentation of two novel methods based on DPM and KCF algorithms.Thus,an intelligent video monitoring system is designed and implemented.The main contents works are as follows:1.A pedestrian detection algorithm extends DPM idea with grid density clustering is proposed.By generating the candidate detection windows,the detection windows redundancy of DPM is naturally reduced.The main idea is: First,moving target detection is obtained by exploiting the three-frame difference algorithm,and the centroid coordinate is further calculated.Then,the dense area is filtered by proposed G-Cluster algorithm.Finally,the candidate detection windows are extracted by the EdgeBoxes algorithm,and the target in which is further detected with classifier.The experimental results show that the proposed algorithm can effectively improve the pedestrian detection consequents of the model.2.A robust KCF algorithm for pedestrian tracking is proposed to effectively deal with tracking drift and tracking failure caused by factors such as scale changes and target occlusion.First,feature fusion is performed by the response distribution of the HOG feature and the HSV feature.Then,a dynamic selection scale pool is set to improve the fixed size matching circumstances of the filter.Finally,the occlusion of the target is measured by the change maximum response rate of the filter.And based on the target information of the last successful frame,EdgeBoxes and the perceptual hash algorithm are used to retrieve the target and update the filter.Experiments show that the proposed algorithm can effectively improve the tracking accuracy and can complete tracking tasks in complex scenes.3.An intelligent video surveillance system is designed and implemented.The algorithms proposed above are integrated into the core functional modules of the system.In addition,it also includes video input module,storage module and area intrusion detection module,forming a monitoring system suitable for real scenes.A running test shows that the intelligent video surveillance system designed and implemented in this thesis has higher intelligence and effectiveness,can better guarantee the security of the surveillance area,and has extremely high application value and prospect.