Tracking Multiple People Under Occlusion Using MultipleCameras

Abstract

We describe a system for tracking multiple people with multiple cameras based on fusion of multiple cues. Face trackers are used to self-calibrate our system. Epipolar geometry and landmarks are employed to disambiguate the tracking problem. The correlation of visual information between different cameras is learnt using Support Vector Regression and Hierarchical Principal Component Analysis to estimate the subject appearance across cameras. The joint features of subjects extracted from multiple cameras are tracked and used as a model to re-track people once the subjects are lost tracking in the system. Results demonstrate that our system can deal with the occlusion.

Ground-Target Tracking in Multiple Cameras Using Collaborative Particle Filters and Principal Axis-Based Integration

“…As no single feature is reliable enough for tracking in all cameras, the fusion of multiple features in the framework of Bayesian Networks was introduced by Chang, et al 6) and Dockstader, et al 8) . The first publication used Bayesian Networks to group features such as color, landmarks, location, and apparent height into targets, while the second tracked 2D semantic features in each camera and fused them by computing the confidence level of a camera using a Bayesian Network.…”

Indoor UAV Control Using Multi-Camera Visual Feedback

“…Since the two same color makers are detected in each camera, the association of measurement (color marker) for UAV tracking is required. This problem is referred to as the data association and has extensively studied in the target tracking and surveillance community [8][9][10]. A number of data association techniques have been developed such as nearest neighbor, the track-splitting filter, joint-likelihood integer programming, multiple-hypothesis algorithm and the joint-probabilistic data association algorithm [11].…”

Probabilistic learning of salient patterns across spatially separated, uncalibrated views

“…The registration of a number of ground-based cameras to the global map allows tracking to be performed across spatially separated camera scenes. The self-calibrated multi-camera system demonstrated in [3] assumes partially overlapping cameras. Epipolar geometry, landmarks as well as a target's visual appearance are used to facilitate the tracking of multiple targets across cameras and to resolve occlusions.…”