DVS Data Recognition Using the Hybrid Neural Network

Since the output of Dynamic Visual Sensor (DVS) is time-continuous event stream data, it has higher temporal order compared to traditional image data, and the event data possesses high sparsity. Therefore, how to effectively extract features from DVS data and recognize specific targets is a challenging task. In this paper, a hybrid neural network approach incorporating Kolmogorov-Arnold Networks and Residual Neural Networks is proposed for extracting effective features from DVS data and performing the task of target recognition. By designing an adaptive convolutional kernel and a dynamic grid adjustment method, the temporal and local features in DVS data are effectively captured. The experimental results show that compared with the traditional Convolutional Neural Network, the proposed method more outperforms in terms of classification accuracy and generalization ability.