A 3.6 μ s latency asynchronous frame-free event-driven dynamic-vision-sensor

1 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX A 3.6µs Latency Asynchronous Frame-Free Event-Driven Dynamic-Vision-Sensor J. A. Leñero-Bardallo, T. Serrano-Gotarredona, and B. Linares-Barranco Abstract − This paper presents a 128x128 dynamic vision sensor. Each pixel detects temporal changes in the local illumination. A minimum illumination temporal contrast of 10% can be detected. A compact preamplification stage has been introduced that allows to improve the minimum detectable contrast over previous designs, while at the same time reducing the pixel area by 1/3. The pixel responds to illumination changes in less than 3.6µs. The ability of the sensor to capture very fast moving objects, rotating at 10K revolutions per second, has been verified experimentally. A frame-based sensor capable to achieve this, would require at least 100K frames per second. I. INTRODUCTION Conventional image sensors are frame-based. In frame-based imagers the detected photocurrent is integrated in a capacitor during a fixed time period (the frame time). The reached voltage level of each pixel is communicated in a sequential way out of the chip. Frame-based imagers have some advantages such as very compact pixels, high fill factor, low fixed pattern noise (FPN), among others. However, they have serious drawbacks such as waste of bandwidth, because all the pixels send out their information regardless of wether they have new information to transmit or not. Also, because photocurrent is integrated over a fixed time period (generally in the order of 20-30ms), information about higher speed moving objects is lost. When a fast moving scene has to be sensed one solution is to reduce the frame period but this generates an overwhelming amount of data to be transmitted and processed. Mechanisms for detection of regions of interest can be applied but this has also high computational costs and delays [1]-[3]. Biological vision sensors operate in a quite different way. When the activity level of a retina pixel reaches some threshold, the pixel sends a spike to its connected neurons. That way, information is sent out and processed continuously in time (in a frameless way) and communication bandwidth is used only by active pixels. Highly active pixels send spikes faster and more frequently than less active ones. Event driven or Address-Event-Representation (AER) [4]-[6] bioinspired vision sensors have become very attractive in recent years because of their fast sensing capability, reduced information throughput and efficient insensor processing. A large variety of AER vision sensors have recently appeared in the literature, such as simple luminance to frequency transformation sensors [7], time-to-first spike coding sensors [8]-[11], foveated sensors [12]-[13], temporal contrast vision sensors [14]-[19], motion sensing and computation systems [20]-[22], and spatial contrast sensors [17]-[18], [23]-[25], just to mention a few. In this paper we present a very low latency AER-based temporal contrast vision sensor. The detection of temporal contrast at the focal-plane level can be very useful to sense and process high-speed moving objects while reducing redundancy and thus maintaining a low level of data to be processed. Several prior frame-based temporal difference detector imagers have been published [26]-[30], however they suffer from limited speed response because they operate based on photocurrent integration during consecutive frames and computing the difference between them. Several event-based (frame-free) temporal contrast vision sensors have been reported in recent years [14]-[19]. They are also referred to as Dynamic Vision Sensors (DVS). The sensor published by Kramer [16] had low contrast sensitivity, while the one by Zaghloul [17]-[18] suffered from poor FPN (fixed pattern 2 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX noise). Lichtsteiner et al. [14]-[15] presented the first practical DVS by introducing a self-clocked switched capacitor differencing and amplification stage resulting in low FPN (2.5%), practical contrast sensitivity (15%), reasonable pixel array size (128x128), very good latency (15µs), excellent intrascene dynamic range (120dB), and sufficient maximum event rate throughput (1Meps). The sensor was appropriate for high speed vision, since it could follow rotating objects up to a speed of 200 rps. Recently, a single pixel for a DVS sensor that achieves a 0.3% contrast sensitivity has been published [31]. The higher contrast sensitivity is achieved by a two-stage differencing amplifier [32], thus increasing the pixel gain a factor 50. However, this new design has very reduced pixel bandwidth. This paper presents a new AER transient vision sensor (DVS) which is based on the one reported by Lichtsteiner and Delbrück [14]-[15]. By introducing a small area non-switched preampliflying stage we obtain a slightly better contrast sensitivity (10%) while reducing the pixel area by a 1/3 factor. By using an alternative photo sensing stage, latency can be reduced down to 3.6µs. As a result, the sensor is capable of detecting rotating objects up to a speed of 10000 rps. The price paid is an increase in the current consumption due to the preamplification stage and a smaller intrascene dynamic range. FPN also increases slightly but it remains significantly lower than the achieved contrast sensitivity. A mechanism has been designed to dynamically adapt the DC levels of the pixel preamplifying stage according to ambient illumination. Thanks to this adaptation mechanism the retina achieves a dynamic range higher than 100dB. However, intrascene dynamic range is reduced to 54dB. II. PIXEL DESIGN The pixel presented in this paper is based on a previously reported design [14]-[15]. However, in the present design a preamplification stage has been introduced that allows to achieve a slightly improved pixel sensitivity while reducing pixel area. Fig. 1(a) shows a conceptual block diagram of the pixel. The first stage, the photoreceptor stage, which is based on a gatebiased source-driven stage [34], converts photocurrent to voltage , (1) where and are the substhreshold slope factor and the substhreshold current factor of transistor [33]. In this type of source-driven stages, gate voltage sets the DC level of the source voltage of transistor for a given photocurrent . The second stage, which is the new preamplification stage, is a voltage gain stage producing an output voltage (2) being the voltage gain of this preamplification stage and a DC voltage. Voltage feeds a capacitive differentiator stage (as proposed in [14]) such that . (3) The capacitive ratio is designed to produce additional voltage amplification. Thus, total voltage amplification from to is . Voltage is compared to thresholds and . When voltage exceeds the upper threshold , a negative event is generated through channel and node is reset to an intermediate reference voltage . When voltage decreases below the lower threshold , a positive event is generated through channel and node is reset to reference voltage . I ph V ph V ph V G nU T I ph I sp ------ ln+= n I sp M p V G V ph M p I ph V pa V off GnU T I ph I sp ------ ln+= G V off V pa dV diff C 1 C 2 ------dV pa –GC 1 C 2 ------nU T dI ph I sp ------ ln– GC 1 C 2 ------nU T dI ph I ph ----------–= = = C 1 C 2 ⁄ V ph V diff A GC 1 C 2 ⁄= V diff V θ + V θ − V diff V θ − OFF V diff V REF V diff V θ + ON V diff V REF 3 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Let us call the voltage excursion at node that generates a single positive event through the ON channel, where is the DC offset voltage of comparator A on in Fig. 1(a), and represents the DC offset voltage of amplifier A d plus the systematic offset voltage introduced at its input node by switching. Similarly, is the voltage excursion at node that generates a single negative event through the OFF channel, with being the DC offset voltage of comparator A off . Referring to the input, let us call “ON contrast threshold” (or “ON contrast sensitivity”) the minimum contrast stimulus that generates a single positive event through output channel ON, and “OFF contrast threshold” (or “OFF contrast sensitivity”) the minimum contrast stimulus that generates a single negative event through output channel OFF. By integrating equation (3), (4) That way, the minimum contrast stimulus to be detected can be adjusted through pixel threshold voltages and . For amplifiers A d , A on , and A off we used simple 7-transistor differential input structures, as opposed to the very simple structures used by Lichtsteiner [14]. The reason was to be able to freely adjust voltage V REF , and also to allow testing of large differences between V REF and . In Section IV.A., we will illustrate how these thresholds must be set to optimize the pixel contrast sensitivity. Next, we describe in more detail the photoreceptor and preamplification stages. A. Photoreceptor The schematic of the photoreceptor block is depicted in Fig. 1(b). An n-well/p-substrate photodiode generates a photocurrent which is converted into voltage by transistor . The photodiode node is clamped to virtual ground, thus increasing the photoreceptor bandwidth and compensating the effect of the large photodiode parasitic capacitance . This input clamping stage [34] avoids any Miller coupling capacitance between the input and output nodes of the voltage amplifier, thus achieving very high bandwidth. Another interesting feature of this source-driven active feedback stage, as analyzed carefully elsewhere [34], is that it becomes stable for operating currents below a top value. The “inverting voltage amplifier” in Fig. 1(b) is always biased with a sufficiently high current (in the order of nano amperes) to not limit the delay of the photo diode branch. Thanks to the avoidance of any Miller capacitance between the amplifier input and output nodes, the sensor’s latency could be improved by a factor of 5 with respect to its prior design [14]-[15]. For low values of the photocurrent , the photoreceptor has a first order dynamics with a dominant pole due to the photodiode node with an associated time constant [34] , (5) where is the transconductance of the voltage amplifier transistor , and and are the output conductances of transistors and respectively (see Fig. 1b). For larger values of the photocurrent , two complex conjugate poles may appear, dominated by the output node dynamics, with an associated time constant and Q V diffON V θ + V REF –( ) V os V osn +( )+= V diff V osn V os V diffOFF V θ − V REF –( ) V osf V os –( )+= V diff V osf θ ev + θ ev − θ ev + I bright I dark -------------- ln V diffON nU T A ---------------- V θ + V REF –( ) V os V osn +( )+ nU T A -----------------------------------------------------------------= = = θ ev − I dark I brigth -------------- ln– V diffOFF nU T A ------------------- V θ − V REF –( ) V os V osf –( )+ nU T A ---------------------------------------------------------------= = = V θ + V θ − V θ +/− I ph V ph M p C d I ph I b I ph τ d g dsb g dsc + g mn -------------------------nU T C d I ph ----------------= g mn M n g dsb g dsc M c M b I ph 4 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX , (6) where is the equivalent capacitance at output node and .Transconductance is the only light dependent parameter changing Q. The worst case Q (maximum) is achieved for , resulting in . Transconductance can be one or two orders of magnitude larger than , but is easily at least two orders of magnitude larger than . Therefore, maximum Q would at the most approach unity. Consequently, in this circuit, even if complex poles appear because of large photo currents, these poles will induce negligible ringing. In Section IV we show that for illumination values up to 30klux, the photoreceptor is the stage that limits the pixel´s bandwidth and the dominant pole is approximately linear with the photoreceptor current as predicted by eq (5). B. Small-Area Preamplifying Stage The schematic of the small-area preamplifier stage is shown in Fig. 1(c). It is composed of two inverting gain stages formed by transistors and , respectively. Each gain stage is preceded by a buffering stage which helps to adapt the DC voltage level at the input of the amplifier. The inverting gain stages are designed to operate in the strong inversion regime, providing a total gain (7) The preamplifying stage was designed to provide a total voltage gain approximately equal to 25. The transistors aspect ratios , , , were designed to operate in strong inversion while at same time having a moderate current biasing in the order of 1µA per pixel. The first stage is designed with larger transistors as its mismatch is amplified by the voltage gain of the second stage. In weak inversion the gain of the preamplifying stage becomes , (8) where and are the subthreshold slope factors [33] of the PMOS and NMOS transistors respectively. So the gain of the preamplifying stage would be reduced approximately to 1, if biased in weak inversion. As we will see in the next Section, voltages and adapt dynamically to global illumination changes, so that global illumination can change over several decades while the preamplifier stages remain properly biased. ndwidth and the dominant pole is approximately linear with the photoreceptor current as predicted by eq (5). C. Differentiator and Comparator Stages The next stage is a self-clocked switched capacitor differencing and amplification stage as used by Lichtsteiner et al. [14]. Lichtsteiner used a gain of 20 for this stage, which required a capacitor ratio of 20, resulting in a high capacitor area consumption. Actually, most of the pixel area (almost 50%) was consumed by these two capacitors. In our case, since we introduce extra gain through the preamplifiers, we only implemented a gain of 5 for the switched capacitor differential stage, thus helping to reduce the overall pixel area. If the CMOS process offers the possibility of using MiM capacitors, Lichsteiner’s design τ o 2C o g dsb g dsc I ph nU T ----------+ + ------------------------------------------= Qg mp g mn C o C d g mp g o +( ) -------------------------------= C o V ph g o g dsb g dsc += g mp g mp g o = Q g mn C o 4g o C d ⁄= g mn g o C d C o M n 1 M p 1 , M n 2 M p 2 , G β n W L -----     n 1 β p W L -----     p 1 -------------------- β n W L -----     n 2 β p W L -----     p 2 --------------------= W L -----     n 1 2,4 µ m 2,4 µ m ----------------= W L -----     p 1 1,2 µ m 16 µ m ----------------= W L -----     n 2 1,2 µ m 1,2 µ m ----------------= W L -----     p 2 0,6 µ m 8 µ m ----------------= Gn p n n -----     2 = n p n n V adc 1 V adc 2 C 1 C 2 ⁄ 5 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX would be much more area efficient. However, even in this case, introducing these small area pre-amplifiers improves the gain (and therefore the contrast sensitivity) by an extra multiplicative factor. For the voltage comparators and the in-pixel AER communication circuits, we used the same circuit already reported by Serrano et al. [35]. III. SYSTEM DESIGN The complete system architecture block diagram is shown in Fig. 2(a). The system is composed of a 128x128 pixel array, a row of 128 preamplifier biasing cells that detect the average illumination along that row which adapts the DC levels of the preamplifiers voltage biases and , a set of programmable current sources that allow to fine tune all the current biases needed by the pixel by accessing to a very reduced set of pins [36], and finally the row and column address event communication circuitry that generates the output addresses. The AER read-out scheme implemented in this design is Boahen’s row parallel technique which latches all the events generated simultaneously in a row and reads them out sequentially speeding up significantly the read out process when high event rates have to be managed [37]. The schematic of a preamplifier biasing cell is detailed in Fig. 2(b). Each cell contains a replica of the pixel photoreceptor circuit that senses the local illumination . The sensed current is replicated by the current mirror formed by the two identical transistors and and summed and low pass filtered in a common node , where a large distributed capacitor eliminates the high frequency components of the sensed current. The result is a spatio-temporal average of the sensed illumination along the row which is locally replicated through a subpico ampere current mirror [38]-[39]. The replicated feeds a modified replica of the pixel photoreceptor circuit where the photodiode has been substituted by the subpicoampere current mirror output. The output voltage of this stage is the input to a replica of the preamplifying stage, where two feedback amplifiers A f1 and A f2 have been added to keep the DC levels of the preamplifiers outputs tied to the external voltage references and . These reference voltages are adjusted to optimize gain and power consumption of the preamplifier stage. Feedback amplifiers A fi (i = 1,2) are detailed in Fig. 2(c). The output transistors M pxi and M nxi are replicated once per pixel array cell (row) plus once per biasing cell (column). This way, the DC current provided to nodes V adci scales directly with pixel array size. The capacitive load at node V xi scales with the number of rows in the array, so that bias current I bf and the compensation capacitor C x would require further readjustments when adding rows. The biasing cells are replicated once per pixel array column to ensure scalability with the number of columns. Nodes , , and are shared by the whole array as well as by the row of 128 biasing cells. Their DC levels are adapted with illumination changes so that the average output voltages of the preamplifiers are always correctly biased in the strong inversion saturation regime. Whenever there is a sudden change of global illumination, this type of retinas produce a sudden storm of output events (also if there is no adaptation mechanism [14]). The time constant of the adaptation mechanism depends on the average photocurrent. We designed it to be in the order of hundreds of milliseconds for illumination levels of about 1klux. IV. EXPERIMENTAL RESULTS A prototype has been fabricated in a double-poly 4-metal 0.35µm CMOS technology. The fabricated retina has a resolution of 128x128 pixels and occupies a total area of 5.5x5.7mm 2 including the pads. Fig. 3(a) shows a microphotograph of the fabricated prototype and Fig. 3(b) a plot of the layout of an arrangement of 4 pixels. Pixels are arranged in a symmetrical way sharing analog and digital routing channels. Thus, noise coupling between analog and digital parts is minimized. Digital lines are routed horizontally over digital parts using metal 2, while sensitive analog lines are routed horizontally over analog parts. Metal 4 is V adc 1 V adc 2 I ph M p M p r V C C C I ph I ph V ph V o1 V o2 V x1 V x2 V adc 1 V adc 2 6 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX reserved to cover the full array except for the photodiode. Table 1 summarizes the main design specifications [40] and compares them with previously reported designs. The retina was tested with a 16mm F/1.4 C-mount lens. Comparing with the design reported by Lichtsteiner et al. [14], this prototype achieves a lower contrast threshold with a 1/3 area reduction. The latency time has also been reduced from 15µs down to 3.6µs. The current consumption increases due to the preamplification stages. FPN also increases slightly but it remains significantly lower than the achieved contrast sensitivity. The retina achieves a dynamic range higher than 100dB. However, intrascene dynamic range is reduced to a factor of 54dB. Sample images are shown in Fig. 4. A. Uniformity of Response, Minimum detectable contrast, and Pixel Gain Characterization. To characterize the uniformity of the response to a given contrast, a similar procedure to the one developed by Lichtsteiner and Delbrück [14] was followed. In order to stimulate all pixels uniformly, a moving gradient bar (shown in Fig. 5(c)) was presented to the retina and the number of positive and negative events generated by each pixel was recorded. The bar crossed the screen in about 4s. For this we used a TFT monitor providing a scene illumination of about 250 lux. The contrast of the stimulus was measured to be (which corresponds to a 1:4 contrast, or 400%). The experiment was repeated for different settings of the pixel thresholds. The bar was swept 30 times for each setting. This way, for each pixel we obtain its corresponding number of positive and negative events fired per edge presentation for the different settings of threshold voltages ( ). Fig. 5(a) shows the histograms for , for both positive (crosses) and negative (dots) events. Histograming the number of events (as done in [14]) does not clearly show the contrast sensitivity of the sensor as a function of threshold voltage settings ( ). However, the same data can be used to histogram , by noting the following. The number of events generated by a pixel for a given stimulus contrast and threshold voltage setting is given by (9) This work Lichtsteiner et al. Gottardi et al. Chi et al. Zaghloul et al. Gruev et al. Posch et al. [19] Resolution 128x128 128x128 128x64 90x90 96x60 189x182 304x240 Fill factor 8.7% 8.1% 20% 17% 14% 30% 10%-20% Latency 3.6µs 15µs 250µs <5ms - 20ms 3µs Consumption 132-231mW 24mW 100µW 4.3mW 63mW 30mW@5V 50mW-175mW Technology 0.35µm 4M 2P 0.35µm 4M 2P 0.35µm 0.5µm 3M 2P 0.35µm 4M 2P 0.5µm 0.18µm 4M 2P MIM Pixel area 35x35µm 2 40x40µm 2 26x26µm 2 25x25µm 2 34x40µm 2 25x25µm 2 20x20µm 2 Chip area 5.5x5.6mm 2 6x6.3mm 2 4.5x2.5mm 2 3x3mm 2 3.5x3.3mm 2 - 9.9x8.2mm 2 Contrast Sensitivity 10% 15% 10% 2.2% - - 13% (single pixel) FPN 4.0% contrast 2.1% contrast - 0.5% of scale 1-2 decades 0.6% of scale - DR >100dB 120dB 100dB 51dB 50dB - 125dB Intrascene DR 56dB 120dB 100dB 51dB - 125dB Table 1. Main design specifications and comparison with previous designs θI bright I dark -------------- ln 1,39= = x y,( ) N + x y,( ) N − x y,( ) V REF V θ +/− – N +/− x y,( ) V REF V θ +/− – θ ev +/− θ N +/− θ θ ev +/− ---------= 7 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Consequently, for each pixel we can compute its contrast sensitivity , and generate the corresponding histogram for as shown in Fig. 5(b). As can be observed, it was possible to set an average contrast threshold as low as with a FPN of σ (θ ev ) = 4.0%. It was possible to adjust a lower threshold value, but then the output was dominated by noise events. This minimum detectable contrast has been improved with respect to the 15% minimum contrast threshold reported in [14] thanks to the increase in pixel gain. It is limited by noise, but also by the mismatch of the comparators (A off and A on in Fig. 1). Tuning too close to the reset voltage of node V diff in Fig. 1, results in some pixels having some of the comparators always ON, and thus self-oscillating. Increasing gain helps in attenuating the effect of this comparator offset mismatch (see eq. (4)). Improving minimum detectable contrast has been precisely the motivation of the present design, increasing the pixel gain to increase the pixel sensitivity, and doing it with a compact circuitry that allows to simultaneously reduce pixel area. Although this preamplifying circuitry introduces some amount of gain mismatch, the 4% FPN remains lower than the 10% overall contrast sensitivity limit. Using eq. (4) and the measured mean values of the contrast thresholds and we can estimate the achieved value of the average pixel gain . By representing the average contrast threshold ( and ) as a function of the voltage threshold , and fitting these graphs to straight lines, we can obtain the average gain from the slopes, as well as the average offset voltages for and . The resulting average gain is , while the offsets are and . Assuming results in and . This latter offset is dominated by reset switching induced offset. B. Dynamic Range The operation of the retina has been verified for over 100dBs of illumination change. The retina has been tested for bright illumination (higher than 50Klux) down to illuminations below 1lux. Maximum ambient illumination was set by using two 20W compact fluorescent light (CFL) bulbs each providing about 30Klux when placed 40cm apart from the retina (scene illumination was measured with an RS 180-7133 digital lightmeter located at the retina position). Then neutral density filters were used to progressively darken the field of view of the retina. The density filters were Newport FS-ND broadband ones that guarantee flat transmission from ultraviolet to near infrared. The lower illumination limit of the retina is limited by the dark current of the photodiodes which is lower than 12fA. We were not able to measure the photodiode dark current with very high precision in our chip, but could establish that it was lower than 12fA. This was done by collecting the total current flowing through node V gnd01a in Fig. 2(b) while keeping the retina in darkness. The current measured through that node in these conditions is 128xI dark plus the leakage current due to the pads and wiring. The total current measured in the darkness through node V gnd01a was 2.33pA. We measured a leakage current for the pads and wiring of 0.76pA, what gives an estimation of the dark current I dark of 12fA. Intra-scene illumination range of the retina is limited, since the preamplifying stage self-adapts to one single global illumination level. Nevertheless, an intra-scene illumination dynamic range of up to 54dB has been verified. Fig. 6 shows the response of the retina with half of the visual field at 1lux while the other half is illuminated at 500lux, by inserting a neutral density filter diagonally over the field of view while placing the setup in a dark chamber to avoid reflections. C. Pixel Bandwidth To measure the pixel bandwidth a group of pixels located in the center of the array were stimulated with a flashing Kingbright super bright red LED L-793SRC-B (1400mcd@20mA) following a similar procedure as the one developed by Lichtsteiner and Delbruck [14]. We created a scene illumination by placing two CFL bulbs near the retina but not directly focused on it. A scene θ ev +/− x y,( ) θN +/− x y,( )⁄= θ ev +/− θ ev 10,45%= V θ +/− A θ ev + θ ev − A θ ev + θ ev − V REF V θ +/− – A V os V osn + V os V osf – A60.8= V os V osn + 103mV= V os V osf – 89mV= V osn V osf = V osn V osf 7mV= = V os 96mV= 8 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX illumination of 30klux was measured by placing the light meter RS-180-7133 in the retina position. The LED was placed in front of the retina. The LED diode was modulated with a sinusoidal signal (10) We varied the frequency of the sinusoid and counted the number of positive and negative events generated by each stimulated pixel per period of the sinewave. The measurements were averaged over a 20 seconds time period. If the magnitude of the stimulus contrast and the voltage thresholds are kept constant, the pixel gain is directly related to the number of generated events (see eqs. (4) and (9)). The measurements were repeated for different values of the illumination by inserting neutral density filters of different attenuation values. These neutral density filters were placed just in front of the retina, thus attenuating simultaneously the scene and the LED illuminations. Fig. 7(a) plots the measured values of the events per cycle as a function of the frequency of the sinusoidal signal for different illumination values. Each of these curves was fitted to a first order transfer function. Fig. 7(a) also shows the fitted first order function corresponding to the lowest illumination of 1.5lux. The optimally extracted poles are plotted in Fig. 7(b) as a function of the illumination. Also, superimposed as a solid trace, the optimal linear fitting of the poles position versus the illumination is shown. As can be observed, in all the measured illumination range the bandwidth is approximately linear with the photocurrent as predicted by eq.(5). D. Latency Latency is the delay it takes from the occurrence of an illumination change at the photodiode until the corresponding output event is transmitted off-chip. To measure the latency, we used the super bright LED stimulated with a step signal [14]. We measured the latency as the delay between the step in the LED signal and the first output request corresponding to an event with the address of the stimulated pixel. In the same way as we did in the previous experiment, we created an intense scene illumination by placing two CFL bulbs near the retina but not directly focused on it. A scene illumination of 30klux was measured by placing the light meter RS-180-7133 in the retina position. The LED was placed in front of the retina. Neutral density filters were placed just in front of the retina, thus attenuating simultaneously the scene and the LED illuminations. We performed the measurements for different values of the illumination. Each measurement was repeated a total of 30 times. The results of these measurements are shown in Fig. 7(c) which plots the measured latency as a function of the illumination. The latency deviation measured among the 30 trials is also marked with error bars. As can be observed, the latency is inversely proportional to the illumination. For very low illumination (in the order of 1 lux) the measured latency is , while for high level of illumination (10 Klux) the measured latency was 3.6µs. The only part of the pixel whose latency depends on illumination is the photoreceptor stage in Fig. 1(a). The rest of circuitry introduces a light independent delay, but depending on biasing. E. Noise Characterizations To characterize noise, we collected noise events (without stimulus) for different settings of illumination and threshold settings. Events are separated according to their sign. Fig. 8(a) represents as a function of the number of noise events per pixel per second collected for each sign, under an ambient illumination of 3 lux. Fig. 8(b) represents the noise events as a function of illumination for , which is approximately the bias we used to determine the minimum contrast threshold in Fig. 5. Noise events are generated at a constant average rate, for a given light and threshold setting. If a stimulus is moving fast, then the ratio of signal to noise events will be high. However, as stimuli move slower, the noise events start to become relevant and the overall signal to noise ratio decreases. V t( ) V off A2πftsin+= 10ms V REF V θ +/− – V REF V θ +/− – V REF V θ +/− – 175mV= θ ev +/− 9 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX F. Power Consumption Chip power consumption has been characterized. The measurements are shown in Fig. 9, where the total chip current consumption is plotted versus the output event rate. The total current consumption includes the current consumed by the programmable current biasing block, the pixel analog and digital parts, and the peripheral AER communication circuitry. At low output event rates the current consumption is dominated by the analog parts, which depends on the chip biasing, while at high event rates the current is consumed mainly by the digital event generation parts. The current consumption at moderate output event rates (below 1Meps) measured in Fig. 9 is approximately 44mA from a 3.3V power supply, that is a consumption of 145mW. Most of this current consumption comes from the strong inversion biased preamplifier stages. For very high event rates (above 1Meps) there is a sudden increase in power consumption because now the contribution of the digital output pads (and all the AER communication circuitry) becomes larger than that of the preamplifier stages. G. Stimulus Rotation Limit This sensor is specially appropriate for very fast moving objects. According to our measurements the sensor can detect moving objects at frequencies higher than 10KHz. We have set up several experiments to generate very high speed moving stimuli. We painted a white dot on the rotating black blade of an electrical fan. The maximum rotating speed of the fan was 400Hz (24Krpm). The retina could easily detect it. Fig. 10 shows a spatio-temporal representation of the positive and negative events captured during a time of 7ms. The gray dots correspond to the positive events and the black dots correspond to the negative signed ones. In order to generate faster and controlled stimuli we used a Tektronix 2467-B analog oscilloscope working in XY mode. We applied as X and Y inputs two signals with a controlled frequency generating a spiral-like trace on the oscilloscope display. However, there was a persistency of the ON state of the oscilloscope phosphor display that made not possible to check the negative transitions at frequencies higher than 1KHz. We could check that positive edges are correctly detected up to frequencies slightly higher than 10KHz. Fig. 11(a) plots the input stimuli waveforms applied to the X and Y inputs of the oscilloscope and Fig. 11(b) shows the image observed in the oscilloscope display for a 500Hz frequency of the input signals captured by a commercial photograph camera. Fig. 11(c)-(d) plot the spatio-temporal representation of the positive and negative events generated by the retina when it is stimulated by the oscilloscope display for different frequencies of the applied signals. Positive events are represented as gray dots while negative events are plotted as black dots. At a stimulus frequency of 500Hz both positive and negative events are clearly distinguished for the spiral moving generated dot. However, for frequencies in the order of 1KHz the negative events begin to become sparse. The problem was related with the stimulus, as we observed that the oscilloscope phosphor display had a persistency time of the ON state that prevented to observe a quick ON-OFF transition. On the contrary, the fast OFFON transition of the oscilloscope display allowed us to observe the ON events of the fast moving dot up to a frequency of 10KHz. The chip illuminance for this experiment was 250lux which is the light provided by the oscilloscope (when setting trace intensity to maximum), measured with our light meter. As shown in Fig. 11(b), the spiral has 5 cycles. Consequently, at 10KHz spiral rotation, a pixel detects a change after 5 cycles, requiring a bandwidth of about 2KHz. According to Fig. 7(a), at this illumination level, a pixel can still provide output for 2KHz. V. CONCLUSIONS A 128x128 temporal contrast retina has been implemented. The new design including a low area preamplifying stage allows to improve contrast sensitivity over previous designs while reducing the pixel area by a 1/3 factor. A global adaptation mechanism 16 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 4. 17 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX (a) (b) Figure 5 VdiffON = VdiffOFF = 0.51V VdiffON = VdiffOFF = 0.44V VdiffON = VdiffOFF = 0.39V VdiffON = VdiffOFF = 0.33V VdiffON = VdiffOFF = 0.28V VdiffON = VdiffOFF = 0.20V 18 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 5 (c) 19 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX 1 lux 500 lux Figure 6 20 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 7(a) 21 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 7(b) 22 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 7(c) Latency (s) 23 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 8 (a) (b) 24 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 9 25 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. XX, NO. X, XXX Figure 10