Multi-Person Tracking in Cluttered Environments Using Hybrid Data Association

Multi-target tracking remains a fundamental problem in robotics, where data association plays a critical role. Algorithms such as the Global Nearest Neighbor (GNN) and the Joint Probabilistic Data Association (JPDA) offer tradeoffs between speed and robustness. GNN is computationally efficient but degrades in cluttered scenes, whereas JPDA enhances robustness through probabilistic reasoning at the expense of higher computation. This paper proposes a hybrid data association strategy that combines GNN and JPDA in a rule-based framework by adaptively switching between them according to scene ambiguity. The method is evaluated on simulated pedestrian-tracking scenarios in ROS and Gazebo with clutter and sensor models with noise. Performance is assessed using T-GOSPA metric and runtime analysis. Results show that the hybrid approach achieves accuracy comparable to JPDA while reducing the runtime gap between JPDA and GNN by about 18.5%, enabling real-time deployment on robotic systems with limited computational resources.