Вы находитесь на странице: 1из 6

Event-based Encoding from Digital Magnetic

Compass and Ultrasonic Distance Sensor for


Navigation in Mobile Systems
Juan Barrios-Avilés, Taras Iakymchuk, Alfredo Rosado-Muñoz, Jose V. Francés-Villora,
Manuel Bataller-Mompeán, Juan F. Guerrero-Martı́nez
Group for Digital Design and Processing (GDDP). ETSE School of Engineering, Dpt. Electronic Engineering
Universitat de Valencia. Av. Universitat s/n. 46100 Burjassot. Valencia. Spain.
Email: Juan.Barrios@uv.es

Abstract—Event-based encoding reduces the amount of gen- event-based theory, from which a certain magnitude value
erated data while keeping relevant information in the measured generates a sequence of events. For maximum efficiency, all
magnitude. While this encoding is mostly associated with spiking three elements of an spiking system, sensors, controller and
neuromorphic systems, it can be used in a broad spectrum
of tasks. The extension of event-based data representation to actuactors, should operate in real time using event-encoded
other sensors would provide advantages related to bandwidth data. Thus, one can avoid the computationally intensive task
reduction, lower computing requirements, increased processing of heterogeneous sensor fusion by using event-based data,
speed and data processing. This work describes two event-based simplifying the controller algorithm significantly.
encoding procedures (magnitude-event and rate-event) for two With the emerging of neuromorphic computing, sensors
sensors widely used in industry, especially for navigation in
mobile systems: digital magnetic compass and ultrasonic distance compatible with the event-based paradigm become an essential
sensor. Encoded data meet Address Event Representation (AER) field of study. As mammal vision is one of the most well
format for further transmission, processing and visualization. studied senses, the vision sensors were one of the first sensors
Two encoding procedures and their associated AER conversion with native event-based information encoding. Starting from
from sensor data are described, using an AVR microprocontroller 1991 with a 50 ×50 pixel basic analog silicon retina [4]
to perform the task in real-time. Results are transmitted to a
computer via USART, displayed and recorded using Matlab or to modern-day Dynamic Vision Sensors (DVS) which are
jAER visualization tool. A comparison with classic transmitting 240× 180 pixel resolution with a dynamic range of 120dB
and processing of sensor data is done to evaluate the convenience and a reaction latency of 12µs [5]. Other well-developed types
of the proposed encoding. Results show that data transmission of event-based sensors are silicon cochleae, which encode the
bandwith can be reduced up to 95% under certain conditions, sound in a similar way to the human ear. Both types of sensors
envisaging that pure or mixed regular/event-based data sensors
are desirable for high speed transmission and low computer were successfully used with software emulating SNN, and
processing while keeping relevant information, which is highly dedicated neuromorphic hardware [6].
desirable for mobile systems. Inertial Measurement Units (IMU) are crucial for robot
navigation and positioning. Usually, an IMU consists of an
I. I NTRODUCTION accelerometer, gyroscope and magnetometer. Accelerometer
Recent progress in the spiking neural network field for and gyroscope provide information about vertical position,
classification, prediction, control, and other applications have movement and vehicle rotation, while magnetic sensor is
lead to an increased interest in neuromorphic hardware. Specif- useful to determine the vehicle orientation relative to the
ically, robotics is one of the fields where most advances external magnetic field (typically, the Earth). The IMU naviga-
are achieved. Spiking controllers were successfully used in tion requires no external beacons, working in every medium:
many practical robotic applications as pencil balancing [1], air, water, underground. In addition, widely used sensors are
ball interception [2], closed-loop toy car race [3]. Potential distance sensors, they are necessary for robot navigation in
benefits of the spiking approach are significant: very low closed spaces and obstacle avoidance. The most popular type
power, robustness, and possibility to use biologically inspired of distance sensors for the medium range (0.1− 10m) are
algorithms. Biologically inspired algorithms are mostly related ultrasonic.
with the recent advances in Spiking Neural Networks (SNN), To demonstrate that different sensors can generate event-
being the last generation of neural networks, they can work based information, we have chosen two different types of
not only with spatial, but also with temporal features of the sensors useful for navigation and mobile systems: a mag-
signal encoded in series of spikes (events), in the same way netic compass providing the magnetic field orientation and
as biological systems process data. However, the critical point an ultrasonic distance sensor. Typical patterns of data output
of current spiking systems is related to the input stimuli and from these sensors include swift rotation, linear movement and
its output. Input and output stimuli must be encoded using oscillations. The aim of this work is to show the possibility

978-1-5090-2870-2/16/$31.00 ©2016 IEEE 640


to include multiple sensor information into spiking systems latency systems; in theory, all event sources are completely
in order to develop fully event-based systems. The proposed independent and asynchronous and generate an AER data
application here is related to mobile robotics but this approach packet immediately after receiving the event. In practical
could be further extended to other sensors according to the applications, simultaneous event collisions are common and
required application. the event source usually includes an event management and
This paper is organized as follows: Section I contains the scheduling algorithm to prevent data loss.
introduction to the problem. Section II describes the hardware Up to the authors’ knowledge, one implementation [9] has
used, how data are read, using AER and visualization. In Sec- been proposed to encode AER from IMU data. In this paper we
tion III, for the magnetic and ultrasonic sensors we describe demonstrate the potential benefits of such data representation.
two event-based encoding methods: magnitude-event and rate- Using easily accessible and inexpensive sensors can be used
event. In Section IV the implementation results are presented as a native event-based sensor. In this case, the ultrasonic
and the comparison of the proposed methods with traditional sensor HC − SR04 and the electronic compass HM C5883L
data generated by sensors is shown. In section V conclusions were used, together with the ATmega2560 microcontroller
are presented. using Inter-Integrated circuit(I2C) for IMU data reading, an
USART for PC interface and two I/O pins for interfacing the
II. M ATERIALS AND METHODS ultrasonic sensor. The HC − SR04 ultrasonic module is able
For this work, we created a test set-up presented in Figure to measure a distance in the range 2−400cm with an accuracy
1. The ultrasonic and magnetometer sensors are connected to that can reach up to 3mm and a maximum update frequency
an AVR microcontroller (ATMega). The microcontroller reads of 75Hz. The magnetic compass uses an HM C5883L IC
the sensors using its standard interface, generates the event magneto-resistive sensor with a 12-bit ADC and 1−2 accuracy.
data based on the magnitude and sends it via a serial port A maximum update frequency of 160Hz can be achieved.
to the PC using Address Event Representation (AER) format.
Once data are received in the PC, two different software tools B. jAER
can be used to read and further process event-based data: One of the tools for visualization of AER dataflow is
Matlab as a standard tool for the mathematical modelling of the open project ’Java Address-Event Representation (jAER)’
the selected methods, and jAER, a widely used tool for event- [10]. It is a Java Graphical User Interface (GUI) tool to display
based systems [7]. A special driver was written to make the and process real-time or recorded event-based data. AER
software compatible with the hardware. devices compatible with jAER include dynamic vision and
Finally, the global system composed by the sensors and the audio sensors, AER monitor/sequencer boards, servo motor
microcontroller act as a neuromorphic sensor providing two controllers and optical flow chips. In addition to data event
event-based encoded values: distance and orientation, being visualization, it is possible to apply or define algorithms and
fully compatible with software tools already developed for filters for event data processing. For compatibility, the AER
other event-based sensors as jAER and Matlab. generated from the microcontroller emulates the ’edvs128’
camera protocol over a 2M Bps serial link for the compat-
A. AER, Address-Event Representation ibility with this software.
This representation was proposed by Sivilotti in 1991 [8]. It
has been widely adopted in the neuromorphic hardware field III. P ROCESSING OF DIGITALLY ACQUIRED DATA:
and the main idea of the encoding is presented in Fig. 2. EVENT -BASED ENCODING
With AER encoding, every device has its own event space This work proposes two encoding algorithms: magnitude-
defined and it only transmits the information about the state event and rate-event. The first algorithm is performed asyn-
changes of its receptors. As an example, for a silicon retina, the chronously, it depends on a change in the magnitude of
event space will be the whole pixel matrix, where every pixel the measurement obtained. Data are generated only when a
is an independent event source. On a threshold crossing event change in magnitude occurs, with respect to a ∆t: the angular
in this pixel, the information about the state change is encoded position with respect to north in case of the compass, and the
in a serial number, or XY coordinates of the changing pixel. distance in case of the ultrasonic sensor, which are translated
Thus, only significant changes generate new data and then, into coordinates and timestamp. The main characteristic of
the amount of information that the retina generates is several this algorithm is that the generation of events significantly
orders of magnitude lower when compared to a traditional reduces transmission bandwidth without losing the relative
vision camera where all data are transmitted continuously. location in case of the ultrasonic sensor and keeping the
In addition, the point of interest (the areas where the image absolute orientation for the magnetic compass, at a very low
changes) is automatically identified since data are generated computational cost.
only from this area. For instance, in a ball intercept task, The rate-event algorithm is built to generate continuous
the average event stream is 20kEvts/s corresponding to a events. It periodically sends the delta time that has elapsed
40kB/s streaming speed; using the classic camera system with since the last event. This information contains the timestamp,
the same time resolution would require 6.6M B/s of raw data resulting in a train pulse encoding an orientation or distance.
stream speed [2]. The AER communication is ideal for low Using this encoding, it is possible to keep the absolute location

641
Fig. 1. Layout of the proposed work: the sensors are read using its data interface, binary data read are encoded into events and AER using a microcontroller
and finally, transmitted to a PC for visualization and processing.

the ultrasonic sensor HC−SR04 is produced. In case of


the magnitude-event algorithms, the current measured value
is compared to the previous value and, if different, an event
is generated. Figure 3 shows the graphical data shown in the
jAER PC software, it can be seen how a moving object in the
range of the sensor is increasing or decreasing position. Four
bytes per event as AER are then transmitted through the serial
port using with the following coding:
1) The first bit of the first byte is used to align the data
Fig. 2. AER encoding scheme. Event-based data from different sensors can (always ’1’) at reception.
be merged in a bus using packets. A REQ-ACK handshake is included to 2) The following seven bits are the Y-axis coordinate.
confirm the event reception and the variable event rate control. Image source: 3) The first bit of the second byte represents the polarity
Mahowald, 1994; Lazarro et al., 1993.
of the event, increment or decrement of the measured
magnitude.
4) The next seven bits represent the X-axis coordinate.
for both the digital compass and the ultrasonic sensor at any 5) Finally, the next two bytes contain the time elapsed since
moment. This procedure is a common type of encoding in the last event generated (timestamp).
biologically-inspired neural systems [11], [12]. The formula of
encoding proposed in this paper is presented in Eq. 1, where
Rmin is an event frequency corresponding to the absence
of stimuli S with Smax the maximum possible value to
be encoded. Note that only positive values can be encoded.
Negative values are encoded as events from a separated event Fig. 3. Ultrasonic data representation in jAER of an object moving forward
source which matches the biological neurons where excitatory and backwards in the ultrasonic distance sensor. White pixels indicate that
and inhibitory synapses exist. The precision and value range the relative distance between the object and the sensor is getting closer, black
pixels indicates the opposite.
of this encoding is limited by the maximum frequency fmax
of events accepted by the source and destination device, and
the temporal resolution ∆t, it is determined according to Eq. In case of the rate-event algorithm, every measured value is
2. The proposed encoding requires a fast transmission channel converted into an event rate (a pulse train). Higher frequencies
with guaranteed time of delivery, or timestamped transmission. imply that a short distance is measured. When coded into
The code for both methods is available as a free source in [13]. four bytes AER, they contain the deltatime(timestamp). The
maximum delta time is 4294967295µs, which is beyond the
|S| ) + R physical measurement limitations of the sensor.
R= f ( min (1)
Smax
Smax B. Event-based and AER with ttY − 271 magnetic compass
fmax (2) sensor.
∆t Once the current orientation value is read from the digital
A. Event-based and AER with HC −
SR04 ultrasonic distance magnetic compass, the arctan(y/x) is calculated so that an
sensor. angle value in radians provides the orientation with respect to
After reading the value from the sensor and estimating the the magnetic polar north. Finally, the radians are transformed
distance, the event generation from the microcontroller for to degrees and adjusted to be only positive.

642
Regarding the magnitude-event encoding, the obtained de- repeatability is guaranteed. Three data transmission methods
grees are transformed into X, Y coordinate, the coordinate is were tested:
also AER coded according to the following protocol: first two 1) Classic data reading and transmission method, i.e. mea-
bytes for alignment, X, Y coordinate and polarity, the next sured values from the standard sensor interface are
four bytes for deltatime. Figure 4 shows the visualization in directly transmitted as 12-bit unsigned binary data using
jAER software: top center position is north and the orien- standard transmission at the maximum update frequency
tation with respect to north is shown. The maximum update allowed by the sensor.
frequency is 160Hz. If there is an event, the deltatime is set to 2) Magnitude-event encoding. After converting read data
7 milliseconds, if not, the delta time is accumulated by 7ms. from sensors into magnitude-events, data are transmitted
only if changes in the measured value arise.
3) Rate-event encoding where continuous event-data are
sent, according to the measured value and procedure
described above.
A. Data transmission rate using the HC − SR04 ultrasonic
sensor distance.
1) Constant measured value.: The first test consisted on
comparing the data sent by the sensor when there is not a
change physically in the distance between the sensor and an
object.
Table I provides the results of encoding at different distances
for an acquisition time of 20s, in this case). The magnitude-
event encoding generates less events and thus, transmission
time is significantly lower than binary data transmission and
thus, more bandwidth is available for other transmission of in-
Fig. 4. Digital magnetic compass data representation in jAER when the formation. The reduction factor is is very high for magnitude-
device rotates on a plane at the same Z-axis. The difference in color for the event since only information for changes is transmitted.
pixels indicates if the device is rotating clockwise or counter-clockwise. Concerning the rate-event enconding, the results in terms
of generated events and transmission rate is the same as
For rate-event encoding, the degree values are encoded in binary data. This is due to the intrinsic encoding method
a pulse train following the encoding proposed in [14] with a where an event data is generated for every measure generating
fixed sample time and 360 addresses where an event is inserted the same number of events and amount of transmitted data.
in the address of the measured value. The update time is 72ms However, data information is represented differently and rate-
in order to have a better implementation of the algorithm as we event encoding is ready to work with spiking systems as
have 72/360 = 200µs as the minimum period per pulse. Next, magnitude-event encoded values.
N events with a delta time fixed at 200µs are generated, where For magnitude-event encoding, the number of events is
N is the degree value measured. Finally, the last delta time for directly affected by the variations in the measured value when
every measure is calculated, corresponding to the difference the system is already stable. For that reason, the number of
between the sample time (72ms) and the result of multiplying events is not directly related to the measured value but to
the degrees (N ) by the fixed period (200µs). the sensor stabilization and noise cancellation. In an ideal
situation the number of events would be one (the changing
IV. RESULTS event) for a constant value. In our tests, the values measured
were converted to events without any further filtering or
The visual results showing the ability of the proposed thresholding in order to obtain the events as raw as possible. If
encoding algorithms to represent the measured value were some mechanical stabilization or low-pass filtering would have
already shown in the previous section. In this section, the main been used, the number of data produced would be significantly
goal is to compare the amount of transmitted data through an smaller.
AER channel, i.e. the data rate when sending event-data from 2) Fluctuating measured values.: For this test, the sensor
the microcontroller to another destination, a PC in this case. was measuring the distance to a point in a camshaft constantly
With this aim, two tests were experimentally developed: the moving (Figure 5). The camshaft movement generated a
first test was done when the measured magnitude is constant constant sinusoidal wave from a perpendicular point of view,
over time, and, in the second case, the measured value is with a period easily manipulated by the rotation speed. Results
fluctuating. are shown in Table I.
Both tests were conducted for a data acquisition time of 20 In this case, all three methods are very similar in the number
seconds, controlling all possible variables of the test such as of generated events since the measured value is constantly
temperature, fixed supports, mechanical variables, etc., so that changing, but it is possible to realize that the gradient of

643
CONSTANT VALUE
BINARY DATA MAGNITUDE-EVENT RATE-EVENT
Distance Total Transm. Total Transm. Reduction Total Transm. Reduction
(cm) Time(s) Events Time (µs) Events Time (µs) Factor % Events Time (µs) Factor %
33 20 1212 6060 212 1060 82.5 1212 6060 0
25 20 1247 6235 882 4410 29.2 1247 6235 0
20 20 1273 6365 599 2995 52.9 1273 6365 0
16 20 1290 6450 374 1870 71.0 1290 6450 0
11 20 1312 6560 725 3625 44.7 1312 6560 0
FLUCTUATING VALUE
2.2 20 1262 6310 1194 5970 5.3 1262 6310 0
2.6 20 1261 6305 1179 5895 6.5 1261 6305 0
3.5 20 1261 6305 1175 5875 6.8 1261 6305 0
5.7 20 1263 6315 1174 5870 7.0 1263 6315 0
15.2 20 1276 6380 1062 5310 16.7 1276 6380 0
TABLE I
ULTRASONIC DISTANCE SENSOR. RESULTS OF MAGNITUDE AND RATE EVENT ENCODING COMPARED TO THE DIRECT SENSOR DATA (BINARY) AT
DIFFERENT CONSTANT AND FLUCTUATING VALUES. T HE TOTAL NUMBER OF EVENTS AND THE ASSOCIATED TRANSMISSION TIME FOR 20s ACQUISITION
IS SHOWN . T HE REDUCTION FACTOR OF EVENT - BASED ENCODING WITH RESPECT TO THE BINARY DATA TRANSMISSION IS ALSO SHOWN .

in Table II.
The reduction factor of magnitude-event encoding compared
to the binary data transmission is higher than that of the
ultrasonic distance sensor (up to 95.1%) since the sensor is
more stable an less sensible to noise. Thus, less events are
generated since no changes in the measured value are detected.
On the contrary, the rate-encoding algorithm produces an
increased number of events since the encoding of the pulse
train requires more data to be transmitted: for every taken
measure, there could be up to 360 ’addreses and delta times’
sent depending on the orientation angle per event, which is
very useful for neural networks training.
2) Fluctuating measured values.: For this test, the compass
was fixed to a rotating blade turning 270 degrees forward and
backward. The results are shown in Table II. As in the previous
test the amount of data generated is significantly reduced
for the magnitude-event encoding algorithm, also being more
stable than in the case of the ultrasonic sensor and thus, less
events generated. The rate-event encoding remains the same
as for binary data, as expected.
V. C ONCLUSION
Applying magnitude-event encoding for the physical sensor
Fig. 5. Experimental setup for fluctuating distance measurement by the data can reduce the amount of data transfer significantly. This
ultrasonic distance sensor, a rotating camshaft provides variable distance effect can be observed for a constant measured value and a
detected by the ultrasonic sensor. Ultrasonic sensor (1), Microcontroller (2),
Camshaft (3). fluctuating movement applied to the sensor as shown in the
results and the visualization in jAER. This not only reduces
the amount of data produced and bandwidth but also allows
the generated data to be directly used to implement algorithm
data produced by the magnitude event has a negative relation and filters already developed in jAER for event-based systems
against the magnitude of the measure. as frameless cameras [15], and spiking systems in general.
Nevertheless, the behaviour and reduction of bandwidth usage
271 magnetic compass
B. Data transmission rate for the ttY − is affected by the quality of the data readings from the sensor
sensor and a filtering or stabilization algorithm could increase the
1) Constant measured value.: In this experiment, the digital reduction of transmitted data.
compass was fixed to a plane parallel to the floor in order The rate-event encoding algorithm proposes an event en-
to correctly sense the orientation in degrees compared to the coding data without reducing data transmission. However,
magnetic north, at any moment. The obtained results are shown it is very useful for post-processing algorithms where this

644
CONSTANT VALUE
BINARY DATA MAGNITUDE-EVENT RATE-EVENT
Distance Total Transm. Total Transm. Reduction Total Transm. Reduction
(cm) Time(s) Events Time (µs) Events Time (µs) Factor % Events Time (µs) Factor %
0 20 2363 11815 116 580 95.0 850680 4253400 ≥ bin
90 20 2363 11815 120 600 94.9 849960 4249800 ≥ bin
180 20 2361 11805 115 575 95.1 850320 4251600 ≥ bin
270 20 2362 11810 116 580 95.0 850680 4253400 ≥ bin
FLUCTUATING VALUE
0 20 2363 11815 208 1040 91.1 4726 23630 ≥ bin
90 20 2362 11810 235 1175 90.0 214942 1074710 ≥ bin
180 20 2362 11810 221 1105 90.6 427522 2137610 ≥ bin
270 20 2362 11810 214 1070 90.9 640102 3200510 ≥ bin
TABLE II
DIGITAL MAGNETIC COMPASS. RESULTS OF MAGNITUDE AND RATE EVENT ENCODING COMPARED TO THE SENSOR DATA (BINARY) AT DIFFERENT
CONSTANT AND FLUCTUATING VALUES . T HE TOTAL NUMBER OF EVENTS AND THE ASSOCIATED TRANSMISSION TIME FOR 20s ACQUISITION IS SHOWN .
T HE REDUCTION FACTOR OF EVENT -BASED ENCODING WITH RESPECT TO THE BINARY DATA TRANSMISSION IS ALSO SHOWN .

data encoding is required as many detection, classification [3] T. Delbruck, M. Pfeiffer, R. Juston, G. Orchard, E. Muggler, A. Linares-
or control event-based algorithms requires this encoding. The Barranco, and M. Tilden, “Human vs. computer slot car racing using
an event and frame-based davis vision sensor,” in Circuits and Systems
results of this algorithm in ultrasonic distance and digital (ISCAS), 2015 IEEE International Symposium on, May 2015, pp. 2409–
magnetic sensors suggest that it could be extended to other 2412.
sensors. [4] M. A. Mahowald and C. Mead, “The silicon retina,” Scientific American,
vol. 264, no. 5, pp. 76–82, May 1991, pMID: 2052936.
Comparing both sensors, the digital compass is less sensitive [5] R. Berner, C. Brandli, M. Yang, S.-C. Liu, and T. Delbruck, A 240x180
to noise in the measure, producing more stability and, then, 10mW 12us latency sparse-output vision sensor for mobile applications.
a better result for both encoding algorithms is generated. IEEE, Jun 2013, pp. C186–187.
[6] V. Chan, S.-C. Liu, and A. van Schaik, “Aer ear: A matched silicon
This result suggests that specific adjustements should be done cochlea pair with address event representation interface,” IEEE Trans-
in event-encoding for each type of sensor. Nevertheless, the actions on Circuits and Systems I: Regular Papers, vol. 54, no. 1, pp.
stability of both sensors could be improved with a simple filter 48–59, Jan 2007.
[7] T. Delbruck, “Frame-free dynamic digital vision,” in Proceedings of Intl.
or thresholding technique. Symp. on Secure-Life Electronics, Advanced Electronics for Quality Life
The event-encoding procedure (magnitude or event) to be and Society, 2008, pp. 21–26.
[8] M. A. Sivilotti, “Wiring considerations in analog vlsi systems, with ap-
used is dependent on the final application. When it is nec- plication to field-programmable networks,” Ph.D. dissertation, Pasadena,
essary to send information at few data, magnitude-event the CA, USA, 1991, uMI Order No. GAX91-37292.
best option, e.g. in an industrial network. However, if it is [9] G. Passetti, F. Corradi, M. Raglianti, D. Zambrano, C. Laschi, and
G. Indiveri, “Implementation of a neuromorphic vestibular sensor with
necessary to produce more data from the same information, analog VLSI neurons,” in Biomedical Circuits and Systems Conference,
as in neuromorphic training the rate-event coding has a better (BioCAS), 2013. IEEE, Oct. 2013, pp. 174–177.
performance and should be used. [10] “jAER Open Source Project,” url: http://jaer.wiki ; sourceforge.net, 2007
[ONLINE].
In summary, this works shows that using event-based en- [11] F. Rieke, D. Warland, R. de Ruyter van Steveninck, and W. Bialek,
coding algorithm requires low computational cost and provides Spikes: Exploring the Neural Code. Cambridge, MA, USA: MIT Press,
event-data ready to be processed with spiking neural networks 1999.
[12] W. Gerstner and W. Kistler, Spiking Neuron Models: An Introduction.
and other neuromorphic systems which can provide efficient New York, NY, USA: Cambridge University Press, 2002.
results in multiple applications. In this case, two sensors [13] J. Barrios, “Ultrasonic and digital compass,” url:
widely used in mobile robotics as distance and magnetic com- https://github.com/juabarav/, 2016 [ONLINE].
[14] A. Linares-Barranco, G. Jimenez-Moreno, B. Linares-Barranco, and
pass are described. The obtained results would allow the use A. Civit-Balcells, “On algorithmic rate-coded aer generation,” Neural
of combined sensor information (vision, distance, orientation, Networks, IEEE Transactions on, vol. 17, no. 3, pp. 771–788, May 2006.
etc.) in neuromorpchic systems so that multiple sources of [15] E. Y. Ahn, J. H. Lee, T. Mullen, and J. Yen, “Dynamic vision sensor
camera based bare hand gesture recognition,” in Computational Intelli-
information can be combined for advanced descision making gence for Multimedia, Signal and Vision Processing (CIMSIVP), 2011
and information processing. All of this at a reduced compu- IEEE Symposium on, April 2011, pp. 52–59.
tational cost, impacting in low power requirements and small
size for hardware, key factors for mobile systems.

REFERENCES
[1] J. Conradt, M. Cook, R. Berner, P. Lichtsteiner, R. Douglas, and
T. Delbruck, A pencil balancing robot using a pair of AER dynamic
vision sensors, May 2009, pp. 781–784.
[2] T. Delbruck and M. Lang, “Robotic goalie with 3 ms reaction time at
4load using event-based dynamic vision sensor,” Frontiers in Neuro-
science, vol. 7, p. 223, 2013, pMID: 24311999 PMCID: PMC3836084.

645

Вам также может понравиться