An Event-Driven Object Recognition Model Using Activated Connected Domain Detection

Address event representation (AER) sensors, recording frameless event data, have recently attracted more attention due to the advantages of sparsed spatiotemporal representation. Spiking neural network (SNN) is a representative biologically plausible model, which is inherently suitable for event-driven AER data processing. However, AER object classification using SNN is still challenging due to the lack of robustness for free moving object recognition. This paper proposes a novel event-driven hierarchical recognition model using an activated connected domain (ACD) location method and an SNN classifier with fusion mechanism. The proposed model extracts bio-inspired cortex-like features by Gabor filters with multiple orientations and scales. Meanwhile, the ACD mechanism coordinates with feature extraction to obtain stable features under random movement of objects. Finally, the features are discriminated by the Tempotron classifier with feature fusion to reduce computing consumption while maintaining comparable performance. Comprehensive experiments conducted on several AER datasets have shown superior performance of the proposed system. Besides, we extend the MNIST-DVS dataset to simulate random moving objects by adding a random continuous spatial offset to the event streams of its samples. Ablation experiments demonstrate that the ACD mechanism enriches the model recognizing capability for free moving objects, especially for training samples only with fixed trajectory movement, which reflects the applicability and robustness. This model equips a high potential for innovation and development in the recognition of moving objects in natural scenes.