Recognition for AER Objects Based on Hierarchical Feature Construction

• Published in: 2022 41st Chinese Control Conference (CCC) pp. 7036 - 7040
• Main Authors: Gao, Tian, Deng, Bin, Wang, Jiang, Wang, Zhiran, Yi, Guosheng
• Format: Conference Proceeding
• Language: English
• Published: Technical Committee on Control Theory, Chinese Association of Automation 25.07.2022
• Subjects: Adaptation models; Address event representation; event-driven; Feature extraction; hierarchical recognition system; Motion segmentation; Neurons; Robustness; spatio-temporal feature; Symbols; Training
• ISSN: 2161-2927
• DOI: 10.23919/CCC55666.2022.9901538

Address event representation (AER) sensors are used to capture dynamic objects ignoring the static redundant information. But how to extract features efficiently from the asynchronous sparse event-driven data with high temporal resolution becomes a problem. In this study, a hierarchical recognition system is proposed to classify AER objects from the event flow. Motion symbol detection (MSD) is applied to adaptively separate the dense event flow into multiple segments with sufficient spatio-temporal information. A hierarchical model composed of two layers is used to construct feature maps based on specified orientations. A convolutional neural network is coupled with the model to further extract feature and classify objects. In the experiment based on MNIST-DVS dataset, it reaches a high accuracy of 99.85% in training and 92.5% in testing.