Speed-based Filtration and DBSCAN of Event-based Camera Data with Neuromorphic Computing

TL;DR

The paper addresses denoising and clustering of high-rate event streams from event-based cameras using neuromorphic hardware. It introduces two hand-crafted spiking neural network architectures: a speed-based filtration network leveraging a neighborhood radius $\epsilon_s$ and speed threshold $t_s$, and a DBSCAN-based clusterer using radius $\epsilon_d$ and min-points $t_d$, both built from integrate-and-fire neurons with fixed weights and delays and without STDP. The authors provide detailed network specifications (general, speed-filter, and DBSCAN variants), analyze per-event runtimes (e.g., $4$ cycles for the speed filter, $t_d + 4$ cycles for DBSCAN border classifications), and discuss hardware compilation strategies to minimize data movement. They validate the approach conceptually on neuromorphic hardware considerations and demonstrate parameter settings (e.g., $t_s=9$, $\epsilon_s=1$, $t_d=10$, $\epsilon_d=3$) with a representative dataset, highlighting the method's scalability and parallelizability for real-time denoising and clustering of event streams.

Abstract

Spiking neural networks are powerful computational elements that pair well with event-based cameras (EBCs). In this work, we present two spiking neural network architectures that process events from EBCs: one that isolates and filters out events based on their speeds, and another that clusters events based on the DBSCAN algorithm.

Figures (5)

• Figure 1: Speed filter network architecture with $\epsilon_{s}$ = 1 that filters out events moving slower than speed_threshold t$_{s}$ = 10.
• Figure 2: Speed filter network architecture with $\epsilon_{s}$ = 1 that filters out events moving faster than speed_threshold t$_{s}$ = 7.
• Figure 3: DBSCAN network architecture with $\epsilon_{d}$ = 1 and threshold t$_{d}$ = 3.
• Figure 4:
• Figure 5: