1

A dataset of sign language words consisted of 11200 samples using DVS sensor is published. In order to further utilize both the spatial and temporal distribution of the events, a new image-like representation of event stream is proposed which consider neighborhood events for frame generating. The proposed frame representation achieves a recognition accuracy of 93.25%.

This work focus on the human gait identification task using event camera, which has much higher temporal resolution and larger dynamic range. A large real-world gait recognition dataset recorded by event camera is released to the community to further facilitate the research on event-based gait recognition. To deal with the unique output of event cameras, a converting method is used to group events into frame forms, thus conventional CNN can be adopted on this representation. To better capture spatial temporal information from events, a new 3D-Graph method is proposed to represent events, and graph neural network is applied to learn to feature embeddding of graph. The proposed graph based approach outperforms CNN-based counterpart by 7.6%.