Address-event Matching for a Silicon Retina Based Stereo Vision System

In this paper we present a stereo vision system based on silicon retina sensors. Derived from the human vision system, these bio-inspired silicon retina sensors are a new type of imagers. Conventional optical sensors capture images at a fixed frame-rate and deliver pixel information of the sensor chip at a certain point in time. In contrast, a silicon retina optical sensor provides only intensity variations (events) in the visual field. The target application of the silicon retina sensor in this work is a stereo vision system for pre-crash/side impact detection. Regarding the asynchronous event-triggered output characteristic of the silicon retina new approaches for data acquisition and processing of the data are necessary. In a first step data from the silicon retina imagers are converted into frames and processed with frame-based stereo vision algorithms. These methods do not benefit from the advantages of the silicon retina technology, thus a new approach exploiting the novelty of the sensor was implemented. This stereo matching approach uses the time information of the delivered address events as a primary matching criterion to find corresponding pixels in the data stream of the left and right sensor. The event-based stereo matching approach is tested with synthetic data, where also an evaluation with ground truth is carried out.