Multi-Camera-Multi-Person Tracking System Based On Tracking-by-Detection Method

Multi-camera tracking has been part of the most exciting applications of computer vision.With the increasing demand of intelligent video surveillance system,single camera system is not sufficient to handle some sophisticated situations such as limited monitoring regions and tracking pedestrians with massive occlusions.In order to address these challenges,this paper proposes a novel framework of multi-camera tracking system which can simultaneously detect and track multiple pedestrians among multiple cameras.The multi-camera tracking brings both opportunities and challenges compared to the single camera pedestrian tracking.It certainly be able to solve the issue of inadequate monitoring regions and massive occlusions problem by providing multiple perspectives and enlarging the views.But doing so involves several main challenges:(1)how to correctly and efficiently share and process visual data collected from multiple cameras for future analysis in the system,(2)how to detect and track different pedestrians individually with distinct appearance and(3)how to recognize and track multiple pedestrians robustly even with massive occlusions.This paper will address these challenges in the following three parts,-Detection: Multi-camera detection is based on single camera detection.The foreground can be obtained by utilizing background subtraction method with multiple fixed cameras in the system.By comparing different approaches of detection,we show that background subtraction with multiple features outperforms other approaches in the experiments,achieves better detection accuracy and is robust to a large variety of backgrounds.-Data association: We can predict the location of pedestrians by calculating the probabilities and applying geometry transformations from different view ranges.Then the locations will be mapped back to each camera to obtain the appearance and location informations for every pedestrian.This strategy empirically shows a superior performance on detection and tracking with massive occlusions.-Multi-camera tracking: The pedestrians can be associated from different videos by introducing a matching score considering both distance to tracking target and appearance model.We present a novel approach which includes features of pedestrians' information like colors and shapes.Then a greedy data association algorithm is designed to seek a higher detection-tracker matching score to map the detections and trackers from different views.When mapping becomes unreliable because of inaccurate location detection,particle filters will be used to assist the mapping process.