A Stable Learning-Based Corner Detector for Event-Based Cameras

Event-based cameras have been raising much attention with their appealing features like high temporal resolution, reduced motion blur, low latency, and high dynamic range. These properties make event cameras more ideal than conventional cameras for computer vision tasks in demanding situations, including in future VR applications. However, event cameras produce an asynchronous stream of signals, which is incompatible with many classic computer vision algorithms relying on frame data. For example, corner detection is one of such algorithms. Corners are low-level and primitive features that can be used in downstream analysis and many applications. The current event-based corner detectors face the difficulty of balancing accuracy, speed, and robustness. In this study, we propose to utilize a machine learning-based method to learn a fast and stable corner detector for event cameras. The experiments show that the proposed corner detector based on Multi-layer Perceptron classifier outperforms current methods in terms of stableness and accuracy, while maintaining a comparable speed. We also propose an enhanced event representation method of time surface, which improves the classifier accuracy and robustness.