Model-driven multi-target tracking in crowd scenes | IEEE Conference Publication | IEEE Xplore

Model-driven multi-target tracking in crowd scenes


Abstract:

Multi-target tracking in crowd scenes is a highly challenging problem due to appearance ambiguity and frequent occlusions between different targets. While many impressive...Show More

Abstract:

Multi-target tracking in crowd scenes is a highly challenging problem due to appearance ambiguity and frequent occlusions between different targets. While many impressive works have been done on complex appearance models and data association framework, we address the importance of social behaviour knowledge to overcome these challenges. The proposed model, termed Crowd Context Model (CCM), offers a general framework which jointly models the appearance features and behaviour rules together, with cooperation methods to achieve model-driven multi-target tracking. We use behaviour modelling approach to make reasonable prediction on pedestrian's location. A Multi-template Appearance Model (MAM) using simple appearance features is adopted for target localization. Experiments on real video sequences show that the proposed model-driven method improves the performance of multi-target tracking successfully, especially during occlusions.
Date of Conference: 09-12 July 2013
Date Added to IEEE Xplore: 21 October 2013
ISBN Information:
Conference Location: Istanbul, Turkey

I. Introduction

Multi-target tracking is an important topic in the field of computer vision. It is a highly challenging problem which aims at extracting trajectory information of targets from video sequences. The problem becomes even more sophisticated when it comes to human tracking, especially in complex and crowded environments where frequent occlusions and interactions would often occur, which makes detection and tracking of targets far more difficult. Figure 1 shows a typical scenario for multi-target tracking in a crowd scene. During recent years detection-based tracking method have achieved impressive progress, which is mostly due to the improvement in object models, either complex appearance models or detectors for specific kinds of objects [5], [6], [7], [8]. The main idea of the method is to link target detection or short tracklets gradually into longer ones, optimizing the global linking scores or probabilities between tracklets. Numerous studies have proved the power of the framework. Current systems are now able to handle long and challenging sequences automatically with high precision.

Contact IEEE to Subscribe

References

References is not available for this document.