Shapes Characterization on Address Event Representation Using Histograms of Oriented Events and an Extended LBP Approach

A new paradigm on visual sensing was introduced in 2006 with the first Event-Driven Dynamic Vision Sensor (DVS) [1], inspired by the asynchronous Address Event Representation (AER) introduced by Mahowald [2], and Kramer’s transient detector concept [3]. The Dynamic Vision Sensor (DVS), also known as silicon-retinae, consists of a 128x128 pixel grid, that captures asynchronous individual light changes at the focal plane, reproducing the behavior of biological retinas [4]. Each pixel on the sensor operates independently and asynchronously, capturing light changes in low latency and high dynamic. When such an event occurs, it is transmitted as an information tuple, indicating pixel position on the grid, time stamp and polarity of the event. Thus this sensor transmits a continuous flow of new events instead of a 2D frame, making it possible to detect event spikes of 10 microseconds (or less), which makes it 100 times faster than high speed conventional cameras (120 fps). This kind of vision sensors is considered as “frameless” providing asynchronous high temporal resolution data. The present paper proposes a methodology to characterize shapes using the flow of DVS asynchronous events based on spatio-temporal features. Fig. 1(a) shows the operation principle diagram of the DVS Address-Event Representation from [4]. Each illumination change on the 128x128 pixel grid generates a spiking event e = ((x, y), t, pol), (x, y) being the coordinates of the pixel on the grid, t the event time stamp, and pol the polarity. Polarity is a binary ON/OFF output. ON polarity informs an illumination increase, and OFF polarity is obtained when illumination decreases, as represented in Fig. 1(a). An event flow composed of N consecutive events is defined as: wN = {e1, ..., eN}. The events of wN can be mapped on a 2D matrix EN , as shown in Figs. 1(b), 1(c) and 1(d),