2024 · 10 citations · 36 references
ComputingAutonomous SystemsComputer ScienceComputer VisionMachine VisionSensor FusionDigital SurveillanceMulti-view GeometryObject RecognitionPose EstimationTracking SystemsVisual TrackingAdaptabilityConsistencyCrowd Management SystemsGeometric Continuity ConstraintsMulti-camera TrackingOnline TrackingPerson Re-identificationTarget Re-identification
Multi-camera multiple people tracking is a crucial technology for surveillance, crowd management, and social behavior analysis, enabling large-scale monitoring and comprehensive understanding of complex scenarios involving multiple individuals across different camera views. However, due to severe occlusion within the scene and significant variations in camera viewpoints, there are high demands for matching and correlating the same target among different cameras, especially in an online setting. To address this challenge, we propose a novel online multi-camera multiple people tracking system. This system integrates geometric-consistent constraints and appearance information of the targets, effectively improving tracking accuracy. Additionally, we design a state-aware Re-ID correction mechanism that adaptively leverages Re-ID features to correct mismatches among targets. This system has demonstrated good adaptability across various scenarios. Our proposed system is evaluated in track1 of the 2024 AI City Challenge [38], achieving a HOTA score of 67.2175% and securing the 2nd position on the leaderboard. The code will be available at: https://github.com/ZhenyuX1E/PoseTrack.
36
Kaiming He, Georgia Gkioxari, Piotr Dollár et al. · 2017 · 28.8K citations
The Hungarian method for the assignment problem
Harold W. Kuhn · Naval Research Logistics Quarterly · 1955 · 12.6K citations
Deep High-Resolution Representation Learning for Human Pose Estimation