Human Activity Recognition with Spiking Neural Network

Most of the research on human activity recognition is now based on deep learning methods due to the rise of deep learning in recent years. As the third generation of artificial neural network, spiking neural network provides an effective bionic solution for nowadays research. Human activity recognition based on event are challenging tasks on computer vision community. At present, most of the event-based datasets are mainly produced by using event camera to record the existing frame-based image datasets, therefore easily contain a lot of noise. In this paper, a spiking neural network with five convolutional leaky integrate and fire layers is proposed to train the event-based dataset. The background activity filtering algorithm is proposed for the event-based datasets with lots of noise, and it is applied to the data augmentation method. Experiments are conduced on the three event-based datasets: DVS-Gesture, IITM-DVS and UCF11-DVS datasets to show that our approach has better accuracy than other methods.