Locate, Track And Re-Build Trajectories Pedestrians In Indoor Space Based On Videos From Multiple Surveillance Cameras

Currently,mature and widely used positioning technology rely on mobile terminals equipped with locators,such as GNSS or GPS,Wi-Fi,and Bluetooth modules.These technologies require activation of the built-in positioning function of the terminal.In some cases,locators may wear dedicated positioning signs,such as Bluetooth tags,RFID tags,etc.However,in open security scenarios,security managers cannot require the jurisdiction of non-specific monitoring targets in the place(such as foreign suspicious individuals)to carry mobile terminals or positioning identification.In such cases,managers often rely on surveillance cameras installed in the premises to analyze the obtained video information manually,to locate and track suspicious targets.However,the efficiency and accuracy of manual positioning and tracking methods can be affected or lowed by complex structures,a large number of cameras,incomplete video coverage,and irregular movement of monitored personnel or targets.To address these issues,this thesis proposes and implements a model and algorithm framework called DeepTrack.This framework is based on multi-source video information,fusion target recognition,and tracking for security application scenarios in large indoor spaces.DeepTrack can achieve automatic indoor personnel positioning,tracking and trajectory reconstruction in multi-storey buildings.The framework involves several models and algorithms:(1)A proposed indoor spatial data model suitable for personnel positioning and tracking that uses point-line networks to express the topological structure and geometric features of indoor space.Two coordinate mapping models are designed to construct the mapping relationship between video pixel coordinates and indoor space coordinates.(2)Integration of the deep learning network YOLO v5 and the multi-target tracking algorithm Deep SORT,to propose a target matching and localization algorithm called Spatial SORT.This algorithm achieves cross-camera target matching based on where the target appears in the overlapping area monitored by multiple cameras.(3)Integration of the target trajectory and shortest path algorithm and space-time linear interpolation obtained by Spatial SORT within the monitoring range.Based on the proposed indoor space data model,complete trajectory estimation and reconstruction of monitored targets in the indoor space is possible.To verify the performance of DeepTrack in practical applications,a multi-storey building was selected as research object in this thesis.We constructed an indoor spatial data model and completed the positioning of surveillance cameras in this model and coordinate mapping of the surveillance range.Five scenarios,including single target,multi-target and cross-floor,were designed to realize the localization and trajectory reconstruction for moving targets in the scenarios.The experimental results show that DeepTrack exhibits high localization accuracy and trajectory reconstruction matching,with an average localization accuracy of 3.81 cm in the case of single target or multiple targets without obstruction.In scenarios with multiple targets and occlusion,the reconstructed trajectory restoration matching degree remains good,but the positioning accuracy declines,and its value is related to the serious occlusion.In the case of occlusion in this experiment,the average positioning accuracy is 29.1 cm.Overall,DeepTrack proposed in this thesis can achieve automatic localization and trajectory reconstruction of moving targets in indoor space.It is expected that this method can help reduce work intensity of security personnel and improve efficiency and quality of security work.