Towards Real-Time Edge Detection for Event Cameras Based on Lifetime and Dynamic Slicing

Abstract

Retinal cameras, such as dynamic vision sensors (DVF), transmit asynchronous events with ultra-low latency ($\sim$10 $\upmu$s) only at significant luminance changes, unlike traditional CMOS cameras which transmit the absolute brightness of all pixels including redundant backgrounds. Due to these significant characteristics, they offer great potential to obtain efficient localization of high-speed and agile platforms. Moreover, event cameras have a high dynamic range (${\sim }$140 dB), which makes them suitable for platforms that operate indoors in low-lighting scenarios and in outdoor environments, where the camera might be pointing at a strong light source, e.g. the sun. In this paper, we propose an algorithm to detect edges in event streams coming from retinal cameras. To do that, an algorithm is developed to extract edges from events by augmenting a batch of events with their lifetimes. The lifetime of each event is computed using a local plane fitting technique. We use a batching technique to increase the frame rate of generated images since events with a high sample rate cause the processing of a single event to be computationally expensive. The size of the batch will be adjusted based on the mean optical flow of the previously generated batch. The obtained experimental results show that our proposed technique can significantly reduce the response time with the same sharpness in generating the edges.