Event-Based Incremental Broad Learning System for Object Classification

Abstract

Event cameras are bio-inspired vision sensors that output pixel-level brightness changes instead of RGB values. Thousands of convolutional neural networks have emerged to process the frame-based images; however, there are few networks designed explicitly for the event-based data, which can fully take advantages of the asynchronous and high temporal resolution data. In this paper, we propose an incremental broad learning system to learn the event-based data in a flat network structure, which consists of feature nodes and enhancement nodes in one layer. The incremental learning strategy is developed for fast adding new nodes in a broad extension, yet it is almost impossible to add a filter or layer in the CNNs without retraining from the beginning. An SVD operation is coupled with the network extension to prevent the redundancy of the network structure. In experiments, our model outperforms the state of the arts, at the same time, 15× faster than the CNNs in training. It makes event cameras easier to be the nearly online training and inference applications.