Event-Driven Dynamic Attention for Multi-Object Tracking on Neuromorphic Hardware

We present a real-time, event-driven multi-object tracking system leveraging the SpiNNaker neuromorphic platform and Dynamic Vision Sensor (DVS). Inspired by hippocampal grid neurons, our approach employs a dynamic attention mechanism based on recurrent spiking neural networks (SNNs) for robust tracking in the presence of distractors. Kalman filtering is used for state estimation and motion prediction, while morphological open/close operations enhance object detection. Our system can track objects even as their motion slows or stops, offering scalable and lowpower multi-object tracking. We demonstrate its effectiveness in dynamic environments, including swarm robot evasion, highlighting its potential for real-time robotics and autonomous systems.