A cascaded timestamp-free event camera image compression method for gesture recognition

Event cameras have drawn great attention due to various advantages compared with frame-based cameras and have been used in many scenarios including gesture recognition. However, due to the high temporal resolution and the existence of timestamp in each event, the huge data amount still poses challenges for real-time processing in household and industrial scenes. To settle this problem, in this paper, we propose a cascaded timestamp-free compression (CTFC) method for reducing redundant data of event camera and demonstrate the effectiveness based on DVS Gesture dataset. The method includes two parts, designed for getting rid of the timestamps and filtering background (BA) noises, respectively. In the first part, a novel event number-fixed compression (ENFC) scheme is presented to remove all of the timestamps directly. The results indicate that $68 \%$ of data are omitted before transmission. In the second part, a spatial-correlation compression (SCC) scheme is proposed to further filter information-independent BA noises. The results show that up to $37.6 \%$ of the residual events are further removed. Moreover, we construct a spiking neural network (SNN) for gesture recognition to monitor the impact of CTFC method on event camera image quality. The simulation results demonstrate that the recognition accuracy is comparable to that of the traditional method after the compression process $\mathbf{(} \mathbf{9 4 . 6 2 \%}$), showing the potential of CTFC method for broadening the real-time application scenarios for event cameras.