`WIDE DYNAMIC RANGE VISION SENSOR
`FOR VEHICLES
`
`Keiichi YAMADA, Tomoaki NAKANO, Shin YAMAMOTO,
`Toyota Central Res. & Develop. Labs., Inc.
`Nagakute, Aichi, 480-11 Japan,
`Eisaku AKUTSU and Keiji AOKI
`Toyota Motor Corp.
`1200 Mishuku. Susono, Shizuoka, 410-11 Japan
`
`Abstract: The dynamic range of brightness on road
`scenes is very wide, because the lighting condi-
`tion dynamically varies with various weather and
`road conditions. Therefore, the dynamic range of
`conventional TV cameras is insufficient to input
`the images of road scenes. We have developed a
`method for expanding the dynamic range of TV
`cameras. Also, we have developed an experimen-
`tal vision sensor system with a wide dynamic range
`based on the method applicable to the vision sys-
`tems for vehicles. The effectiveness of the sensor
`in comparison with conventional TV cameras was
`confirmed from the experiments on highways un-
`der various lighting conditions.
`
`I. INTRODUCTION
`Computer vision greatly contributes to preventive safety
`and/or driver assisting systems for vehicles, because it is
`effective in recognizing lane marks, cars and obstacles on
`the lane [l, 2,3]. The dynamic range of brightness on road
`scenes is very wide where the lighting condition dynam-
`cally varies according t o various weather and road condi-
`tions. One of the problems in realizing the vision systems
`for vehicles is that the image taken with TV camera is
`sometimes either hidden under noise in its dark parts or
`saturated in its bright parts. This is because the dynamic
`range of conventional TV cameras is insufficient to input
`the images of these road scenes [4]. Accordingly, a TV
`camera with a wide dynamic range is necessary to realize
`such a computer vision system for vehicles.
`One of the approaches t o expanding the dynamic range
`of TV cameras is to improve imaging devices [5]. Although
`many studies on imaging devices have been done, no such
`a device that satisfies tbe requirements for vehicle applica-
`tion, a wide dynamic range, enough resolution, sensitivity
`and reasonable cost, has been reported. Another approach
`is to compose an image with a wider dynamic range than
`that of a camera from the images taken under different
`exposure conditions. Some studies on still images with a
`wide dynamic range of the objects at rest have been re-
`
`0-7Ro3-2105-7/94/$4.00
`
`1994 IEEE
`
`ported (6, 71. However, a practical TV camera which can
`take moving object has not yet been realized.
`We have developed a method for expanding the dy-
`namic range of T V cameras which can take moving objects.
`Based on the method applicable to the vision systems for
`vehicles, we have developed an experimental vision sensor
`system with a wide dynamic range. The effectiveness of
`the sensor in comparison with conventional TV cameras
`was confirmed from the experiments on highways under
`various lighting conditions.
`
`11. REQUIRED DYNAMIC RANGE
`FOR A VEHICLE CAMERA
`
`Fig. 1 shows brightness of objects, presenting on road
`scenes in the daytime, under various lighting conditions.
`The figure means that a dynamic range of up to lo4 is re-
`quired for a camera to be capable of taking images of the
`road scenes without lack in information. This indicates
`that the dynamic range of conventional TV cameras, ap-
`proximately 500, is insufficient t o take the images of the
`scenes without the lack.
`Conventional N camera
`+11111111------,
`Requied dynamic range
`
`101
`
`102
`
`103
`
`b
`Brightness
`1041"~)
`
`A s " c
`La"smah
`surface
`o n m d
`Carbody Carbody
`(Gray) While)
`In thetwilight ( 1 0 % ~ )
`
`m r m d Lanemm
`onroad
`sutiace
`Carbody Cart&
`(Gray)
`(Whne)
`In the sun at noon (1051Ux)
`
`Carbody Carbody
`(Gray)
`(write)
`In the tunnel (i021ux)
`
`carbody Carbody
`(whnel
`(Gray)
`In the shade at noon (1041ux)
`
`Fig. 1. Brightness of objects in road scenes under various
`lighting conditions and the dynamic range required for a
`camera to input the scenes.
`
`111. METHOD FOR EXPANDING
`DYNAMIC RANGE
`An image with a dynamic range wider than that taken
`
`1994 Vehicle Navigation & Information Systems Conference F%ceedm
`
`gs
`
`405
`
`Magna 2017
`TRW v. Magna
`IPR2015-00436
`
`
`
`by a TV camera would be obtained by combining the im-
`ages of a scene taken under different exposure conditions.
`To establish the method of expanding the dynamic range
`of a TV camera, assume the following model representing
`the relation between input light intensity S and output
`signal level L of a TV camera. Assume that the output
`signal level L is proportional to the light intensity S until
`L is saturated at the saturation level Laat :
`
`where E is a coefficient decided by the exposure condition.
`Suppose a set of exposure conditions Eo, E l , ..., E,-,, each
`condition is represented by:
`
`In equation (2), A, which is greater than 1, is an increase
`rate of exposure amount.
`Fig. 2 shows the output signal level L versus light
`intensity S in each exposure condition EO, Et, ..., En-].
`The light intensity in each range, expressed by the fol-
`lowing equations, is detected using each image taken with
`exposure condition E, ,
`
`Note that L,,, represents a noise level of the camera. By
`detecting the light intensity in each range from each image,
`as is also shown in Fig. 2, the light intensity range between
`L,,t/EoAn-l and L,,,/Eo would be covered by n images
`taken with exposure conditions EO to En-1.
`Since a conventional TV camera takes images under
`a single exposure condition, the light intensity range ob-
`tained by the conventional camera corresponds to that in
`the exposure condition EO. Let DO be the dynamic range of
`the conventional camera. Then, the dynamic range which
`can be covered by exposure conditions Eo to En-l is ex-
`pressed as DoA"-', which is An-' times as wide as the
`range of the conventional camera.
`
`Camera output level
`1% L t
`
`Lsat
`
`Lnoi
`
`IV. WIDE DYNAMIC RANGE
`VISION SENSOR
`An experimental vision sensor system, whose dynamic
`range is expanded by the method described in the previ-
`ous section, has been developed. Fig. 3 shows the block
`diagram of the sensor. The sensor is composed of a CCD
`monochrome TV camera and a VME double height size
`electronic circuit board named dynamic range ezpansion
`unit. The images in different exposure conditions are taken
`at short intervals by using electronic shutter function of
`the CCD camera to realize the imaging of moving objects.
`The images taken with different exposure time are com-
`bined into an image by the dynamic range expansion unit
`simultaneously with the reading out from CCD: Thus, the
`image with a wide dynamic range could be obtained in real
`time. The appearance of the developed sensor system is
`shown in Fig. 4.
`Fig. 5 shows the timing diagram when the number of
`exposure condition is two. Shutter time is varied in every
`field to take the image in each exposure condition every
`1/60 seconds. Image signal is read out from the CCD
`camera in non-interlace mode. As shown in the figure,
`when the number of exposure condition is two, two images
`for each of these exposure conditions can be taken in 1/60
`seconds if the image for the shorter exposure time is taken
`first: Thus, the wide dynamic range image is obtained
`every 1/30 seconds.
`The dynamic range expansion unit works as follows:
`The image for the first exposure condition is stored in the
`image memory. Each image taken after that is combined
`with the previous image read out from the image memory.
`The combined image is sent back to the image memory and
`stored there. By combining the image for the last exposure
`condition with the image from the image memory, a wide
`dynamic range image is obtained.
`The combining algorithm is as follows: The algorithm
`is processed by one pixel. If the pixel of the image for
`the longer exposure time is not saturated, this pixel value
`is used as the value of the combined image. Otherwise,
`the pixel value for the shorter exposure time is used by
`multiplying it and the ratio of the two exposure times to
`
`Dynamic range expansion unit
`
`Fig. 2. Camera output level L vs. light intensity S in each
`exposure condition E, (where i=1,2, ..., n-1).
`
`Fig.3. Block diagram of developed vision sensor system.
`
`406
`
`1994 Vehide Navigation & Ia60mtion Systems Conferenae Ratedingr
`
`
`
`Fig. 4. Appearance of developed vision sensor system.
`
`-
`
`Time
`
`Exposure term
`(approximately 1/60 sec.)
`IC*
`
`Read out
`
`Combine
`
`output
`
`Fig. 5. Timing diagram of the developed vision sensor
`when the number of exposure condition is two.
`
`of IO4 is obtained. This dynamic range is considered to
`be wide enough for vehicle cameras from Fig. 1.
`'The
`performance of the vision sensor is summarized in Table
`I
`
`V. EXPERIMENT
`A. Experimental Method
`To evaluate the effectiveness of the developed vision sen-
`sor. we performed the following experimcnts under various
`lighting conditions on the road.
`A conventional TV camera and thc camera ticad of the
`developed vision sensor were placed on a r o d of an au-
`tomobile a t a height of l.G m and declination angle of 5
`degrees. The image from the conventional camera was dig-
`itized with a frame grabber into 2.56 gray levels and saved
`for t,he evaluation. The image from the developed vision
`sensor was obtained from the image memory of the dy-
`namic range expansion unit and saved for the evaluation.
`Driving on highways a t a speed of 80 km/h: t,he images
`were taken with the conventional TV camera and devel-
`oped vision sensor simult~aneonsly. The weather was fine.
`The intensity of illuminatior ranged approximately from
`lo2 lux. in a tunnel. to 10' lux, in the sun. The ninii-
`ber of exposure condit.ion of the developed vision sensor
`was two. and exposure lime was set at 1/87 and 1/2620
`seconds. A fixed iris lens was used for t,he developed vi-
`sion sensor; an auto iris lens was used for the conventional
`camera t o follow the change in the brightness of t,hc scenes.
`The obtained
`Focal length of both the lenses was 25".
`images were evaluated in terms of lack in information by
`comparing the two images simultaneously taken with the
`conventional camera and the developed sensor.
`
`i oL
`
`? 256
`
`i o o
`
`io1
`
`i o 2
`
`103
`
`lo4
`
`Light intensity (arbitary unit)
`Fig. 6. Performance of developed vision sensor: pixel
`value of combined image with wide dynamir range vs. in-
`put light intensity when the exposure time is 1/87 and
`1/2620 seconds.
`
`Table 1. Performance of developed vision sensor
`
`Resolution
`Minimum illumination
`Exposure time
`Number of exposure time
`Image memory size
`
`564(H) X 242(V) pixels
`5 lux. E'1.4
`63.6 ps to 1/60 sec.
`2 to 5
`1024(H) x 256(\') x 2 bytes
`
`t When the exposure time is 1/87 and 112620 seconds.
`
`normalize the pixel value t o the scale of longer exposurc
`time pixel value. The algorithm is implemented as a look-
`u p table for speedup.
`To evalnate the dynamic range of t,he vision sensor,
`the gray level of the obtained wide dynamic range image
`in relation to t,he light intensity was measured. Fig. G
`shows the result for the exposure timr of 1 / 8 i and 1 /2G2O
`seconds. As can be seen from the figurc, a dynamic range
`
`B. Experimental Results
`Fig. i ( a ) and 7(b) show the image taken \rit,h the convcn-
`tional T V camera and that with the sensor at, t.h(, exit of
`a tunnel. respectivcly. A t,unnel exit, is one of \,tie scenrs
`having the widest dynamic range exccpt the casc t.hat the
`sun rays directly come int,o t,he lens. A s can he been from
`the figures, the developed vision sensor obtains a clearer
`image of the scene than t,he image wit,h the corivent,ional
`T V camera. On t,he other hand. some parts of the iiii-
`age obtained with the conventional camera are saturated
`and lack in information due t,o the narrow dynamic range.
`Fig. 8(a) and 8 ( b ) show anot.hrr example of the images of
`the scene where the sunlight was very bright and shade\\-s
`were very dark. The effectiveness of 1 tic developed scmor
`can be seen from the figures in t,hc same manncr as Fig. T.
`Although 29 L7c of the pixels in the image in Fig. i ( a ) and
`8 5% of the pixels in Fig. 8(a) are sat,urated, the pixel in
`both the images obtained with the developed serisor (Fig.
`7(b) and 8 ( b ) ) arc ncit,her saturated nor zero Irvcl.
`Fig. 9 ( a ) and 9 (b) show the cdge enharrccmcnt result.
`of the images in Fig. 7 ( a ) and i ( b ) using a sobcl operat,or.
`All of the lane mark edges are clearly observed in t,he image
`
`1994 Vehicle Navigation & Information Systems Conference Proceedings
`
`407
`
`
`
`The dynamic range of a conventional TV camera, 500,
`was expanded to lo4 with the developed image sensor sys-
`tem. The dynamic range of the developed sensor was found
`to be sufficient to input images of the road scenes in vari-
`ous lighting conditions. The effectiveness of the sensor in
`comparison with conventional TV cameras was confirmed
`from the experiments on highways under various lighting
`conditions.
`At present, the experimental vision sensor developed
`is not small enough in dimension for practical application.
`This is because most of the digital logic circuits in the
`dynamic range expansion unit are constructed by general-
`purpose logic ICs. However, it is possible to reduce the
`dimension for practical application. This is because the
`digital circuits can be integrated into a custom LSI. In ad-
`dition, both the analog circuits for video signal and the
`A/D converter, which are necessary for digital image pro-
`cessing for the vision system, commonly exist in conven-
`tional frame grabbers. Consequently, only the logic LSI
`and a look-up table for combining images are the essen-
`tial parts to be added to the conventional frame grabber
`for expanding the dynamic range of a TV camera. There-
`fore, the vision systems for vehicles would easily adopt
`the method for expanding the dynamic range of the TV
`camera and increase the robustness under varies lighting
`conditions.
`
`REFERENCES
`
`[l] E. D. Dickmanns and V. Graefe, “Dynamic monocular
`machine vision,” Machine Vision and Applications, 1,
`pp.223-240, 1988.
`[2] A. Kutami, et al., “Visual navigation of autonomous
`on-road vehicle,” IROS ’90, pp.175-180, 1990.
`[3] T. Ozaki, et al., “Image processing system for au-
`tonomous vehicle,” SPIE(Mobi1e Robots), 1989.
`[4] C. Thorpe, M. H. Hebert, T. Kanade and S. A.
`Shafer, “Vision and navigation for the Carnegie-Mellon
`Nablab,” IEEE Trans. PAMI, vol. 10, no. 3, pp.362-373,
`1988.
`[5] S. G. Chamberlain and J. P. Y. Lee, “A novel wide
`dynamic range silicon photodetector and linear imag-
`ing array,” IEEE Trans. ED., vol. 31, no.2, pp.175182,
`1984.
`[6] R. M. Rangayyan and R. Gordon, “Expanding the dy-
`namic range of x-ray videodensitometry using ordinary
`image digitizing devices,” Appl. Opt., vol. 23, no. 18,
`pp.3117-3120, 1984.
`[7] K. Moriwaki, “Adaptive exposure image input system
`for obtaining high quality color information”, IEICE,
`vol. J76-D-11, no. 9, pp.1894-1901, 1993.
`
`(b)
`( a)
`Fig. 7. Images at a tunnel exit obtained with; (a) a
`conventional TV camera, (b) the developed vision sensor.
`
`Fig. 8. Images on the road where the sunlight is very
`bright and shadows are very dark obtained with; (a) a
`conventional TV camera, (b) the developed vision sensor.
`
`(b)
`(a)
`Fig. 9. Edge enhancement result of the images in Fig.7
`using a sobel operator.
`
`for the developed sensor while some parts of the edges are
`lacked in the image for the conventional camera.
`From the experimental results described above, the ef-
`fectiveness of the method for expanding the dynamic range
`of a camera for the vision systems of vehicles has been con-
`firmed.
`
`VI. CONCLUSION
`
`One of the problems in realizing the vision systems
`for vehicles is insufficient dynamic range of conventional
`TV cameras. To solve the problem, we have developed a
`method for expanding the dynamic range of TV cameras.
`Also, we have developed an experimental vision sensor sys-
`tem with a wide dynamic range based on the method ap-
`plicable to the vision systems for vehicles.
`
`408
`
`1994 Vehiclc Navigation & Informatl ‘on Systcmr Confcrcnae Roeetding