A Less Noisy Time Surface for Event-based Visual Odometry

Event cameras have recently drawn an increasing amount of attentions due to its advantages in high temporal resolution, reduced motion blur and high dynamic range compared to traditional frame-based camera. However, a single event output by the event camera carries limited information, and it is challenging to be applied to fields of robotics such as visual odometry. An advanced method to handle this problem is to introduce time surface (TS) into the event-based stereo visual odometry. TS provides a compact spatio-temporal presentation for events and develops the potential of event cameras. Whereas, TS is sensitive to noises from raw events and the polarity of the event is handled separately in the state-of-art research, which could diminish the performance of TS. Motivated by these problems, this paper proposes a less noisy TS approach to reserve more environmental details under noisy conditions. It is realized with three additional efficient operations on the local memory TS, as well as a weighted combination of positive and negative TS map. Three operations include considering less weighted neighbor information, treating events that do not occur in groups and events of large time intervals as noisy events. Experiments on the open-source dataset are conducted to verify the effectiveness of the proposed less noisy TS by applying it to an event-based visual odometry.