Efficient, event-driven feature extraction and unsupervised object tracking for embedded applications

Neuromorphic vision sensors offer a low-power, bandwidth efficient way to extract salient visual information from the scene and are a candidate for energy-efficient embedded systems. An algorithm for embedded, event-driven feature extraction and object tracking to leverage such sensors is outlined and demonstrated. Near sensor data sparsification and information extraction is conducted in three distinct steps: memory efficient noise filtering, fast scalable identification of keypoints, and subsequent clustering to identify objects in the scene. The processing flow has demonstrated rates of near 100 fold data reduction, 5 fold improvement of feature extraction throughput, and sustenance of an event processing rate of 212kAEps.

Paper

Full text

PDF

Efficient, event-driven feature extraction and unsupervised object tracking for embedded applications

Semantic Scholar · Engineering · 2021

Abstract

Neuromorphic vision sensors offer a low-power, bandwidth efficient way to extract salient visual information from the scene and are a candidate for energy-efficient embedded systems. An algorithm for embedded, event-driven feature extraction and object tracking to leverage such sensors is outlined and demonstrated. Near sensor data sparsification and information extraction is conducted in three distinct steps: memory efficient noise filtering, fast scalable identification of keypoints, and subsequent clustering to identify objects in the scene. The processing flow has demonstrated rates of near 100 fold data reduction, 5 fold improvement of feature extraction throughput, and sustenance of an event processing rate of 212kAEps.

Similar papers

© 2026 NYSGPT2525 LLC