Event-VoxelHop: A Polarity-Wise Successive Subspace Learning (SSL) Algorithm for Event-Based Object Classification

In this study, we introduce Event-VoxelHop, an innovative algorithm designed for object classification using event cameras with successive subspace learning (SSL). Unlike traditional vision systems that rely on frames, event cameras detect changes in light intensity at the pixel level, offering high temporal resolution and low-energy use. We apply an SSL algorithm, a simpler yet effective approach compared to traditional CNN-based methods, retaining the inherent low latency advantage of event cameras. Our method enhances classification accuracy and efficiency by processing event polarity and building upon the SSL framework by generating event volumes from short streams, effectively capturing both spatial and temporal dynamics. This approach significantly improves object classification accuracy with event-based cameras by utilizing both temporal and spatial information. Through extensive testing on the CIFAR10-DVS dataset, we demonstrated a notable improvement in classification performance, achieving an accuracy of 76.3%, substantially higher than existing methods. This achievement clearly shows the method’s effectiveness in providing more accurate and energy-efficient solutions for event camera-based object classification.