[Paper Review] FA-Harris: A Fast and Asynchronous Corner Detector for Event Cameras

Ruoxiang Li, Dianxi Shi+3 more|arXiv (Cornell University)|Jun 26, 2019

Advanced Memory and Neural Computing26 references4 citations

TL;DR

This paper proposes FA-Harris, a fast and asynchronous corner detector for event cameras that operates directly on raw asynchronous events without synthesizing event-frames. It employs a Global Surface of Active Events (G-SAE) maintenance unit and a corner candidate selection mechanism to enable real-time performance, while a refined Harris-based method ensures high accuracy, achieving approximately 8× speedup over prior event-based Harris detectors with no loss in accuracy on public datasets.

ABSTRACT

Recently, the emerging bio-inspired event cameras have demonstrated potentials for a wide range of robotic applications in dynamic environments. In this paper, we propose a novel fast and asynchronous event-based corner detection method which is called FA-Harris. FA-Harris consists of several components, including an event filter, a Global Surface of Active Events (G-SAE) maintaining unit, a corner candidate selecting unit, and a corner candidate refining unit. The proposed G-SAE maintenance algorithm and corner candidate selection algorithm greatly enhance the real-time performance for corner detection, while the corner candidate refinement algorithm maintains the accuracy of performance by using an improved event-based Harris detector. Additionally, FA-Harris does not require artificially synthesized event-frames and can operate on asynchronous events directly. We implement the proposed method in C++ and evaluate it on public Event Camera Datasets. The results show that our method achieves approximately 8x speed-up when compared with previously reported event-based Harris detector, and with no compromise on the accuracy of performance.

Motivation & Objective

• To address the lack of efficient, truly asynchronous corner detection methods for event cameras that operate directly on asynchronous event streams.
• To overcome the limitations of frame-based approaches that rely on artificially synthesized event-frames, which compromise asynchrony and introduce latency.
• To design a real-time corner detection pipeline that maintains high accuracy while processing events at high rates.
• To enable robust feature extraction for SLAM and visual odometry in high-speed and high-dynamic-range environments using event cameras.

Proposed method

• The method processes asynchronous events through an event filter to remove noise and reduce data volume.
• A Global Surface of Active Events (G-SAE) is maintained in real time to represent spatio-temporal event activity, enabling efficient computation.
• A corner candidate selection unit identifies potential corners based on local G-SAE curvature and gradient changes, prioritizing high-response regions.
• A corner candidate refinement unit applies an improved event-based Harris detector to score and validate each candidate using local intensity variations on the G-SAE.
• The entire pipeline is implemented in C++ and operates directly on event streams without requiring intensity frames or frame synthesis.
• The algorithm is optimized for low-latency processing, with computational performance measured per-event processing time and maximum event rate (Meps).

Experimental results

Research questions

• RQ1 Can a corner detection method be designed to operate natively on asynchronous event streams without frame synthesis?
• RQ2 How can real-time performance be achieved in event-based corner detection while preserving accuracy?
• RQ3 What is the impact of G-SAE-based processing on speed and accuracy in corner detection for event cameras?
• RQ4 How does the proposed method compare to existing event-based corner detectors in terms of speed, accuracy, and false positive rate?

Key findings

• FA-Harris achieves approximately 8× speedup over the previously reported eHarris* detector, with a processing time of 0.66 μs per event and a maximum event rate of 1.54 Meps.
• The method maintains high accuracy, achieving a weighted average accuracy of 50.88% across all scenes, outperforming eHarris* (50.56%) and G-eHarris* (50.57%).
• FA-Harris reports a lower False Positive Rate (FPR) than other methods, with a 2.58 FPR on the overall dataset, indicating better precision in corner detection.
• The algorithm demonstrates robustness in dynamic and high-motion scenes, with a 3.28% FPR on the 'dynamic' scene and 1.61% on the 'boxes' scene.
• Despite being 5× slower than the Arc* method, FA-Harris achieves significantly better accuracy and lower FPR, indicating a favorable trade-off between speed and precision.
• The evaluation confirms that FA-Harris operates directly on asynchronous events without frame synthesis, preserving the intrinsic asynchrony of event cameras.