Applications of event‐based image sensors—Review and analysis

Received 26 April 2018; Revised 6 June 2018; Accepted 6 July 2018 DOI: xxx/xxxx RESEARCH ARTICLE Applications of Event-based Image Sensors —Review and Analysis† Juan A. Leñero-Bardallo* | Ricardo Carmona-Galán | Ángel Rodríguez-Vázquez 1Institute of Microelectronics of Seville (IMSE-CNM), Consejo Superior de Investigaciones Científicas y Universidad de Sevilla, C/ Américo Vespucio s/n, 41092, Seville, Spain Correspondence *Juan A. Leñero-Bardallo. Email: [email protected] Summary The spread of event-driven asynchronous vision sensors during the last years has increased significantly the industrial interest and the application scenarios for them. This article reviews the main fields of application that event-based image sensors have found during the last twenty years. We focus in the description of applications where such devices can outperform conventional frame-based sensors. The practical functions of the three main families of asynchronous event-based sensors are analyzed. The article also studies what are the factors that increase nowadays the demand of sensors that minimize the power and bandwidth consumption. Moreover, the technological factors that have facilitated the development of asynchronous sensors are discussed. KEYWORDS: Image sensors, event, spike, Address Event Representation, AER, asynchronous 1 INTRODUCTION Since Mahowald and her colleagues published in the early nineties the first integrated asynchronous bio-inspired vision sensor1, the spread of event-driven devices has grown significantly2. Depending on the information from the visual scene that is processed, there are three main categories of event-based (also known as event-driven or just spiking) vision sensors. The first, and the oldest group, corresponds to sensors that detect spatio-temporal contrast within the visual scene. In their origins, they were bio-inspired sensors that try to emulate the early processing of the visual scene that is performed in the human retina. The second and the most successful category so far corresponds to Dynamic Vision Sensors (DVS). These are devices whose pixels spike whether they detect transient illumination changes. They can track very fast moving objects with very low power and bandwidth data consumption. Finally, the third category corresponds to spiking luminance sensors, also known in the literature as octopus retinas. Although they just perform a light-to-frequency conversion, they preserve the asynchronous communication †This work was supported in part by the Spanish Ministry of Economy and Competitiveness under Grant TEC2015-66878-C3-1-R, Co-Funded by ERDF-FEDER; in part by Junta de Andalucía CEICE under Grant TIC 2012-2338 (SMARTCIS-3D); and in part by ONR under Grant N000141410355 (HCELLVIS). 2JUAN A. LEÑERO-BARDALLO ET AL based on spikes to encode the illumination levels. Only illuminated pixels can send information off-chip. Their dynamic range and their latency are quite competitive. Thus, they are attractive for applications where is necessary to detect illumination levels of certain regions with low resources consumption. Although the commercial success of asynchronous vision sensors is still far away from frame-based sensors, there are several companies like Prophesee or Inilabs dedicated to the fabrication of event-based cameras. The rise of the demand of vision sensors with low power and bandwidth consumption, and the capability of processing the visual scene has put spiking sensors into the scene. Drone vision and autonomous vehicles prefer devices with such features instead ones that provide high image quality. For these reasons, asynchronous image sensors find accommodation and are highly competitive in situations where image quality is not a must. There are several reviews and comparisons between the different event-based sensors published so far2,3. However, to the best of our knowledge, still there is not still published a review of the specific fields of application where they have been employed. In this article, we will focus on a description of a the fields of application where event-based vision sensors have been successfully accommodated. We will pay special attention to applications of the family of spiking luminance sensors in which the authors have made relevant contributions during the last years. 2 SENSORS BASED ON SPATIAL CONTRAST DETECTION Spatial contrast detection is the very first-processing of the visual scene performed by the human visual system4. It is implemented by the cells that compound the retina. Their outputs are transmitted as train of spikes through the optic nerve. The detection of spatio-temporal contrast allows to detect edges, and shapes. That is crucial for an early interpretation and classification of the elements within the scene. At first instance, spiking sensors capable of computing spatial contrast on the focal plane pursue to mimic a biological modelling of the human retina4. The first silicon retina chip was proposed by Mahowald,1. The original pixel topology is shown in Fig. 1 . Pixels detect light with a photoreceptor circuit which output is a voltage that is averaged within a neighborhood with a diffusive resistive network that interconnects all the pixels. If the average illumination value in the neighborhood, 𝑉𝑝𝑛𝑒𝑖𝑔ℎ𝑏. , differs from the pixel’s local illumination value, 𝑉𝑝𝑖,𝑗 , the pixel will spike, indicating the detection of spatial contrast. These circuits, in the literature, still they are known as silicon retinas due to their biological inspiration. There are improved versions of this preliminary approach5,6,7. This family of sensors opened a field of research that lead to ulterior development of event-based vision sensors. They also put emphases in the development of asynchronous circuitry to implement the well-known AER (Address Event Representation) communication protocol8,1,9. Modern arbitration schemes10,11 are based on the early versions implemented on the first asynchronous bio-inspired vision sensors. JUAN A. LEÑERO-BARDALLO ET AL 3 + _ + _ Photoreceptor Right Top Bottom Left C Pixel’s Output R R R R Vpi,j Vpneighb. ___ Vpneighb. ___ FIGURE 1 Retina contrast detection pixel according to1. Average illumination values within a pixel neighborhood (𝑉𝑝𝑛𝑒𝑖𝑔ℎ𝑏. ) are computed with a diffusive resistive network. The average illumination value in the neighborhood (𝑉𝑝𝑛𝑒𝑖𝑔ℎ𝑏. ) is compared to the local pixel illumination (𝑉𝑝𝑖,𝑗 ). If both values differ, the pixel spikes, with a frequency proportional to the spatial contrast magnitude. Once the possibility of implementing these sensors was demonstrated, the community drew its attention about how to improve the quality of their images. Particularly, mismatch was an important challenge associated to the first generation of bio-inspired sensors, operating with transistors in subthreshold region. Several authors pointed out the need of reducing pixel Fixed-Pattern- Noise (FPN) associated to transistor mismatch. In that sense, calibration procedures and circuitry robust to mismatch were developed12,7. Nowadays, the event-driven spatial contrast computation is not highly competitive in terms of spatial noise against the traditional method: read-out an entire pixel matrix illumination values and process them on the digital domain to detect edges and shapes. The main reason is that asynchronous computation of spatial contrast on the focal plane requires non-linear operations that are difficult to implement precisely in the analog domain. However, they deserve to be mentioned in this survey because they led to the development of other families of asynchronous sensors whose applications will be described in the next sections. 4JUAN A. LEÑERO-BARDALLO ET AL + _ Cph V reset + _ + _ -A C 1 C 2 Reset V bot V top OFF EVENT ON EVENT v α log(I ) iph v = −C /C·(dv /dt) diff 1 2 i FIGURE 2 Architecture of a DVS sensor. Light is detected by a photologarithmic receptor with negative feedback and gain −𝐴. Its voltage outputs feed a differencing circuits that amplifies transient voltage variations at its input, 𝑣𝑖. If the differencing circuit output, 𝑣𝑑𝑖𝑓𝑓 , exceeds a negative or a positive voltage threshold, events are transmitted indicating temporal contrast. The sign of the temporal contrast (positive or negative) is detected with two comparators and transmitted through two independent pathways. 3 DVS SENSORS The pixel’s architecture of a Dynamic Vision Sensor (DVS) sensors is depicted in Fig. 2 . Light is processed by a logarithmic photoreceptor. Its output voltage, 𝑣𝑖, feds a differencing circuit that computes and amplifies its temporal derivative, 𝑣𝑑𝑖𝑓𝑓 . If the transient variation of differencing voltage exceeds a programmable threshold (either positive or negative), and event is sent offchip, indicating a pixel’s illumination change, its coordinates, and the sign of the temporal contrast. DVS sensors can detect fast changes of illumination, with low power and data bandwidth consumption. Movement detection is outperformed over transitional procedures, based on frame processing. Proof of that is the fact that several companies like Inilabs or Prophesee have put on their scope the development of these family of sensors. The first functional DVS sensor was presented by Lichtsteiner et al. in 200813. In this first publication, the authors already devised possible applications for the sensor, i.e. traffic monitoring, gaming, sleep monitoring, etc. In more recent publications, several authors presented improved versions of the original DVS sensor14,15,16,17. Regarding the application fields of these sensors, we can highlight the same examples: 1. Datasets for implementation and training of spiking neural networks for object tracking, action recognition and/or object recognition. Spiking neural networks needs to be fed with spiking inputs. Images generated with asynchronous image sensors are ideal for this purpose. The spread and popularization of DVS sensors among the community has driven the creation of event-based image datasets recorded with DVS sensors18,19. JUAN A. LEÑERO-BARDALLO ET AL 5 2. Traffic monitoring. The control of moving objects is an ideal task for DVS sensors. Only pixels that detect illumination changes send data off-chip, minimizing the bandwidth and the power consumption. There are specific papers addressing this topic, for instance,20. Also DVS sensors have been used to implement proximity detectors,21. 3. Fall detectors. Processing the output data flow of a DVS sensor, the authors propose a fall detector22. It was intended for homecare of elder people. In can be integrated among the sensors of an intelligent house. Gesture recognition is also possible,23. 4. Robotics integration, control and navigation24. The idea is to provide images of visual scenes with cameras that are mounted on systems that are continuously moving or observing a non-static visual scene. 5. Tracking and particles detection on fluids. The excellent temporal resolution of DVS sensors has made possible to track fast movements of particles in fluids reducing the data throughout25. 6. Color change detection26. Combining the pixel architecture of a DVS sensor with stacked photodiodes at different depths in standard silicon, it was possible to create an image sensor whose pixels only reacts to color changes (transient shifts within the spectrum). 7. Stereo vision27,28. The combination of two or more DVS sensors recording images from different localizations allows determine distances and depth, after processing their outputs. The DVS sensor reduced output data flow allows to alleviate the bandwidth and power processing consumption with regards to classic frame-based processing algorithms. 4 SPIKING LUMINANCE SENSORS In this paper, we will focus on summarizing practical applications where asynchronous spiking sensors have been employed. These devices, also known as octopus retinas, perform a light-to-frequency conversion to encode luminance levels within the visualscene29.Theiroperation principle is depicted in Fig. 3 .Pixeloperationis initiatedbyresetingthevoltage attheintegration capacitance, 𝐶. Then, its voltage decreases linearly with a slope that is proportional to the pixel’s photocurrent until it reaches a programmable analog voltage, 𝑉𝑟𝑒𝑓 . Thereafter, the pixel self-resets and starts integrating charge again. The pixel’s oscillation period is approximately given by this expression: 𝑓𝑜𝑠𝑐 ≈𝐼𝑝ℎ 𝐶⋅(𝑉𝐷𝐷 −𝑉𝑟𝑒𝑓 )=𝐼𝑝ℎ 𝐶⋅Δ𝑉(1) Where 𝐼𝑝ℎ is the pixel photocurrent and 𝐶is the integration capacitance. 𝑉𝑟𝑒𝑓 is a voltage threshold to control the pixel sensitivity to light. Although these sensors do not implement any kind of processing, just measuring luminance levels, they have inherent advantages over conventional frame-based image sensors. The first one is speed; pixel spiking frequency can 6JUAN A. LEÑERO-BARDALLO ET AL c V (t) t spike 1 spike 2 spike 3 M1 D1 Iph V c V c Vref Vref V DD -I /C ph + _ C VDD spike VDD Vref _req_x reset_x _req_y _ack_y reset_y AER Communication Circuitry ∆V FIGURE 3 Integrate-and-fire (I&F) pixel block diagram. Pixels spike with a frequency proportional to light intensity. Every time that this occurs, pixels’ addresses are transmitted asynchronously off-chip. Asynchronous digital signals involved in the data transmission are displayed on the right. easily reach KHz rates in indoor environments30,29. This implies an equivalent temporal resolution in the order of Kilo-frames per second. The second advantage is their autonomous operation. Pixels have not to be scanned periodically to gauge their illumination values. They pulse continuously whenever they are illuminated. If they are dark, they do not send any data. That is also an advantage for application scenarios where dark pixels are meaningless. Another strength of them is their High Dynamic Range (HDR) operation. Pixel spiking frequencies can spam more than five decades31,29. This implies the capacity of detecting intra-scene illumination levels with a dynamic range higher than 100dB. Also, their power consumption is reduced with respect frame-based sensors. Their main penalty is image quality. FPN noise is typically higher with these sensors30,31. However, there many applications that can sacrifice image quality in benefit of speed, HDR operation, and low power consumption. We summarize and explain some of them in the next subsections. 4.1 Sun Sensors Sun sensors detect the sun position (azimuth and latitude) referred to their centroid. They are demanded for green energy production in solar plants and to develop navigation systems for spacecrafts and sounding rockets, that take the sun as a reference for orientation. There are two main classic families of sun sensors: analog33 and digital34,35. Analog ones are more simple and JUAN A. LEÑERO-BARDALLO ET AL 7 θ x y FD D T Pixel Matrix Lid & Pinhole Rmax φ (a) (b) 10010110210 310 4 Degrees (º) 72.5 72.6 72.7 72.8 72.9 Latitude (θ) Error (º) 0 0.1 0.2 0.3 0.4 θ (º) Error (º) Number of Spikes (N ) 10010110210 310 4 Degrees (º) 34 34.5 35 35.5 36 36.5 Azimuth (φ ) Error (º) 0 0.5 1 1.5 2 2.5 φ (º) Error (º) Ν =1 spikes Ν =1 spikes spikes (c) FIGURE 4 (a) Operation of a sun sensor32 built with a pinhole lens. (b) Snapshot of an asynchronous sun sensor implemented with a pinhole lens. (c) Sensor’s output data. The sun position is calculated computing the centroid of a certain number of incoming events. With only one event, it is possible to gauge the sun position with an error lower than 2.2%. easy to design, but they are more prone to noise and sensitive to scene distractors. Digital ones are the most extended ones nowadays. They are reliable and easy to build. Deploying a dedicated optics on top of a frame-based sensor, it is possible to convert it into a sun sensor. Depending on the sun position, a group of pixels will be illuminated. Its centroid will depend on the sun localization. By processing the output data, the centroid is calculated. Then, establishing basic trigonometric relations, the sun position can be determined. The main limitations of this approach are the data redundancy and the limited operation speed. Data redundancy is an important drawback because dark pixels needs to be readout and are useless to compute the sun position. This also drives to high bandwidth and power consumption derived of the data readout and the data processing. The second limitation is the operation speed. Since the entire pixel matrix has to be readout, every time that the sun position has to be updated, the operation speed is conditioned by the frame rate. Several authors have proposed strategies to avoid reading dark pixels34. However, still some of them have to be readout periodically. A recent alternative to build digital sun sensors is the usage of octopus retinas36,32. By placing a dedicated optic over them, only pixels exposed to sunlight will send data. Therefore, data redundancy is avoided, simplifying the data processing, the sensor operation and reducing power consumption. Moreover, the excellent temporal resolution of octopus retinas make these devices suitable for space navigation. In Fig. 4 .a, we display the architecture of a spiking sun sensor with a pinhole lens on top of it. In Fig. 4 .b, the fabricated spiking sun sensor is displayed. Data recorded with the sensor is provided in Fig. 4 .c. The centroid computation of the illuminated region is fast and simple. Just computing the centroid of the pixels that fires, it is possible to do that. The operator has freedom to decide how many spikes are employed to compute it. Usually the centroid is closer to the highest 8JUAN A. LEÑERO-BARDALLO ET AL 20 40 60 80 20 40 60 80 100 120 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Secondary Combustion zone Interzonal Region Primary Combustion Zone Relative NIR levels 4 5 6 7 8 9 10 Time (s) 100 150 200 250 300 Relative Energy in the NIR band Transient Variations of NIR energy in the Scene -100 -80 -60 -40 -20 0 20 40 60 80 100 Frequency (Hz) 40 50 60 70 80 90 |X(jw)| (dB) (a) (b) FIGURE 5 Flames monitoring with an asynchronous NIR spiking sensor38,39. (a) Snapshot of a flame taken with a luminance asynchronous sensor with an optical NIR filter. (b) Top: Transient NIR levels variations detected with the sensor in a visual scene with a flame. Bottom: FFT of the previous data to study the flame frequency components. illuminated pixel that spikes first than the others. Thus, just taking into account its coordinates, it is possible to gauge the sun position with acceptable precision. In terms of dynamic range, spiking sensors allow to encode the illumination levels without saturating. Usually, frame-based sensors dynamic range is limited to 70dB37, that is not enough to detect illumination levels inside very bright light sources like the sun that require sensors with more than 100dB dynamic range to be characterized37. 4.2 Flame monitoring Asynchronous spiking sensors can be used in the detection and the analysis of flames activity,38,39. Flame flickering and flame emissions have to be monitored in some industrial processes40,41. These tasks were implemented, at first instance, with infrared (IR) cameras or microbolometers. Their main limitations are the high cost, the limited frame rate, the need of calibration in some cases, and their fragility42. As alternative to them, CMOS or CCD cameras working in the Near Infrared Band (NIR) have been employed. By placing a NIR filter over them, the become NIR imagers. Radiation inside this band is emitted by flames. Hence, with adequate image processing, flames activity can be monitored43,44,45. Flames flicker with frequency components that range between DC and 100Hz40. Thus, the analysis of their transient activity requires sensors with good temporal resolution. This requirement also JUAN A. LEÑERO-BARDALLO ET AL 9 P- N- P- N+ I&F I&F I&F IB IGB IRG RB GRGB (a) (b) FIGURE 6 Asynchronous color sensor implementation,46. (a) Stacked photodiodes to achieve spectral selectivity within the visible spectrum. Each one is connected to integrate-and-fire (I&F) block as the one depicted in Fig. 4 . (b) Input stimulus and color images rendered with an asynchronous sensor implemented with the proposed pixel architecture. implies that, operating with a frame-based camera, the amount of data to be processed is be very high, limiting the data analysis to off-line operation40. Recently, the authors advanced how octopus retinas can used to monitor flame activity,39. The excellent temporal resolution of these sensors can be exploited to track their transient activity efficiently. Filtering the radiation below the NIR band, the output data is limited to the regions exposed to NIR radiation. Hence, their output data flow can be processed real-time39, outperforming the operation obtained with classic NIR imagers40. In Fig. 5 .a, a snapshot of a flame taken with an octopus asynchronous sensor fabricated in standard CMOS technology is displayed. An optical filter was placed over the sensor to remove the radiation outside the NIR band. Results are alike the ones that can be obtained with a commercial infrared camera. It is possible to identify the main regions inside the flame. On the top of the right plot in Fig. 5 .b, we have plotted the transient NIR levels detected by the sensor in a visual scene with flames. Flames provoke fast variations of the NIR levels. On the bottom of the same plot, the Fast Fourier Transform (FFT) of the previous data is displayed. Main frequency components range from DC to 30Hz, for this flame. 4.3 High speed color detection with stacked photodiodes Currently, standard CMOS technologies offer the possibility of incorporating a deep n-well or a deep p-well to the chip design. These layers allow to stack photodiodes at different depths, placing a deep n-well beneath an n-well, as it is depicted in Fig. 6 .a. Stacked photodiodes have different sensitivity to light wavelength, depending on their depth47,48. Hence, similar spectral discrimination can be achieved than employing traditional methods with discrete optical filters. In this previous work46, there diodes were stacked to encode color at high speed. In Fig. 6 .a, there is a sketch of the color pixels. Each one is connected to an integrate-and-fire (I&F) module with the same architecture than an octopus sensor as the one shown in Fig. 3 . Processing the 16 JUAN A. LEÑERO-BARDALLO ET AL 24. Delbruck T., Villanueva V., Longinotti L.. Integration of dynamic vision sensor with inertial measurement unit for electronically stabilized event-based vision. In: 2014 IEEE International Symposium on Circuits and Systems (ISCAS):2636-2639; 2014. 25. Drazen David, Lichtsteiner Patrick, Häfliger Philipp, Delbrück Tobi, Jensen Atle. Toward real-time particle tracking using an event-based dynamic vision sensor. Experiments in Fluids. 2011;51(5):1465. 26. Farian L., Leñero-Bardallo J.A., Häfliger P.. A Bio-Inspired AER Temporal Tri-Color Differentiator Pixel Array. Biomedical Circuits and Systems, IEEE Transactions on. 2015;9(5):686-698. 27. Domínguez-Morales Manuel Jesus, Cerezuela-Escudero Elena, Perez-Peña Fernando, Jimenez-Fernandez Angel, Linares- Barranco Alejandro, Jimenez-Moreno Gabriel. On the AER Stereo-Vision Processing: A Spike Approach to Epipolar Matching. In: Neural Information Processing:267–275Springer Berlin Heidelberg; 2013; Berlin, Heidelberg. 28. Domínguez-Morales M., Jimenez-Fernandez A., Paz R., et al. An Approach to Distance Estimation with Stereo Vision Using Address-Event-Representation. In: Neural Information Processing:190–198Springer Berlin Heidelberg; 2011; Berlin, Heidelberg. 29. Culurciello E., Etienne-Cummings R., Boahen K.A.. A biomorphic digital image sensor. Solid-State Circuits, IEEE Journal of. 2003;38(2):281-294. 30. Leñero-Bardallo Juan A., Carmona-Galán R., Rodríguez-Vázquez A.. A wide linear dynamic range image sensor based on asynchronous self-reset and tagging of saturation events. IEEE Journal of Solid-State Circuits. 2017;52(6):1605-1617. 31. Leñero-Bardallo Juan A., Carmona-Galán R., Rodríguez-Vázquez. A.. A bio-inspired vision sensor with dual operation and readout modes. Sensors Journal, IEEE. 2016;16(2):1-14. 32. Leñero-Bardallo J. A., Farian L., Guerrero-Rodríguez J. M., Carmona-Galán R., Rodríguez-Vázquez Á.. Sun Sensor Based on a Luminance Spiking Pixel Array. IEEE Sensors Journal. 2017;17(20):6578-6588. 33. Ortega P., Lopez-Rodriguez G., Ricart J., et al. A Miniaturized Two Axis Sun Sensor for Attitude Control of Nano-Satellites. IEEE Sensors Journal. 2010;10(10):1623-1632. 34. Xie N., Theuwissen A. J. P.. A Miniaturized Micro-Digital Sun Sensor by Means of Low-Power Low-Noise CMOS Imager. IEEE Sensors Journal. 2014;14(1):96-103. 35. Boldrini F., Monnini E., Procopio D., et al. Attitude sensors on a chip: Feasibility study and breadboarding activities. In: Proceedings 32nd Annual AAS Guided Control Conference:1197âĂŞ1216; 2009. JUAN A. LEÑERO-BARDALLO ET AL 17 36. Farian L., Häfliger P., Leñero-Bardallo J. A.. A Miniaturized Two-Axis Ultra Low Latency and Low-Power Sun Sensor for Attitude Determination of Micro Space Probes. IEEE Transactions on Circuits and Systems I: Regular Papers. 2018;65(5):1543-1554. 37. Darmont Arnaud. Methods to extend the dynamic range of snapshot active pixel sensors. In: Proceedings of SPIE Vol. 6816, 681603 (2008); 2008. 38. Leñero-Bardallo J. A., Guerrero-Rodríguez J. M., Carmona-Galán R., Rodríguez-Vázquez Á.. On the analysis and detection of flames with an asynchronous spiking image sensor. IEEE Sensors Journal. 2018;:1-9. 39. Leñero-Bardallo J.A., Bryn D.H., Häfliger P.. Flame monitoring with an AER color vision sensor. In: Circuits and Systems (ISCAS), 2013 IEEE International Symposium on:2404-2407; 2013. 40. Yan Yong, M. Colechin, R. Hill. Monitoring of oscillatory characteristics of pulverized coal flames through image processing and spectral analysis. IEEE Transactions on Instrumentation and Measurement. 2006;55:226-231. 41. Jones A R. Flame failure detection and modern boilers. Journal of Physics E: Scientific Instruments. 1988;21(10):921. 42. Briz S., Castro A.J., Aranda J.M., Melendez J., Lopez F.. Reduction of false alarm rate in automatic forest fire infrared surveillance systems. Remote Sensing of Environment. 2003;:19-29. 43. Cheon Jimin, Lee Jeonghwan, Lee Inhee, Chae Youngcheol, Yoo Youngsin, Han Gunhee. A Single-Chip CMOS Smoke and Temperature Sensor for an Intelligent Fire Detector. IEEE Sensors Journal. 2009;9:625-628. 44. Maoult Y. Le, Sentenac T., Orteu J.J., Arcens J. P.. A New Approach Based on a Low Cost CCD Camera in the Near Infrared. Fire Detection. 2007;:193-206. 45. Bendiscio P., J F. Francescon, Malcovat P., Malobert F., Polett M., Valacca R.. A CMOS integrated infrared radiation detector for flame monitoring. ISCAS. 1998;6:625-628. 46. Leñero-Bardallo Juan A., Bryn D.H., Häfliger P.. Bio-Inspired Asynchronous Pixel Event Tricolor Vision Sensor. Biomedical Circuits and Systems, IEEE Transactions on. 2014;8(3):345-357. 47. Lange R.. 3D Time-of-Flight distance measurement with custom solid-state image sensors in CMOS/CCD-technology. PhD thesisDepartment of Electrical Engineering and Computer Science at University of Siegen2000. 48. Sze S.M.. Semiconductor Devices Physics and Technology. AT&T Bell Laboratories: Wiley; 1985. 49. Leñero-Bardallo Juan A., Carmona-Galán R., Rodríguez-Vázquez Á.. Live Demonstration HDRLVS. In: https://www.youtube.com/watch?v=KrdpUpBRD60. 18 JUAN A. LEÑERO-BARDALLO ET AL 50. Leñero-Bardallo Juan A., Häfliger P.. A Dual-Operation-Mode Bio-Inspired Pixel. Circuits and Systems II: Express Briefs, IEEE Transactions on. 2014;61(11):855-859. 51. Ruedi P.-F., Heim P., Kaess F., et al. A 128 ×128 pixel 120dB dynamic-range vision-sensor chip for image contrast and orientation extraction. Solid-State Circuits, IEEE Journal of. 2003;38(12):2325-2333. 52. Barbaro M., Burgi P. Y., Mortara A., Nussbaum P., Heitger F.. A 100×100 pixel silicon retina for gradient extraction with steering filter capabilities and temporal output coding. IEEE Journal of Solid-State Circuits. 2002;37(2):160-172. 53. Shoushun Chen, Bermak A.. Arbitrated Time-to-First Spike CMOS Image Sensor With On-Chip Histogram Equalization. Very Large Scale Integration (VLSI) Systems, IEEE Transactions on. 2007;15(3):346-357. 54. Son B., Suh Y., Kim S., et al. 4.1 A 640×480 dynamic vision sensor with a 9µm pixel and 300Meps address-event representation. In: 2017 IEEE International Solid-State Circuits Conference (ISSCC):66-67; 2017. 55. Leñero-Bardallo J. A., Rodríguez-Vázquez Á.. Analog Electronics for Radiation Detection. In: Analog Electronics for Radiation Detection. 2016 (pp. 47-67). JUAN A. LEÑERO-BARDALLO ET AL 19 TABLE 1 Comparative between different spiking sensor families Type of sensor/ Functionality Author / Year / Key first implementation Key parameters Key applications Spatial contrast detection Mahowald 1991,1Spatial contrast sensitivity Bio-inspired modelling, applications with reduced output data flow, surveillance, early scene interpretation DVS sensors Lichtsteiner 2008,13 Temporal contrast sensitivity Datasets for spiking neural networks, traffic monitoring, robotic integration, domotic systems, stereo vision Octopus sensors Culurciello 2003,29 Dynamic range and latency Light-to-frequency conversion, sun sensors, flame monitoring, fast color encoding Hybrid sensors Several authors 2010,16,14,30 Dynamic range, latency, power consumption, spatio-temporal contrast detection Combinations of features from different sensors families, HDR imagers, motion detection