Multi-Resolution Vehicle Detection using
`Artificial Vision
`
`Pier Claudio Antonello
`Alberto Broggi, Pietro Cerri
`Dipartimento di Ingegneria dell’Informazione Surround Sensing - Vehicle development
`Universit`a di Parma
`Centro Ricerche FIAT
`Parma, I-43100, Italy
`Orbassano (TO), I-10043, Italy
`{broggi, cerri}@ce.unipr.it
`pierclaudio.antonello@crf.it
`
`Abstract
`This paper describes a vehicle detection system using a
`single camera.
`It is based on the search for areas with a high vertical
`symmetry in multi-resolution images; symmetry is computed
`using different sized boxes centered on all the columns of
`the interest areas.
`All
`the columns with high symmetry are analyzed to
`get the width of detected objects. Horizontal edges are
`examined to find the base of the vehicle in the individuated
`area. The aim is to find horizontal lines located below an
`area with sufficient amount of edges. The algorithm deletes
`all the bounding boxes which are too large, too small, or
`too far from the camera in order to decrease the number of
`false positives.
`All the results found in different interest areas are mixed
`together and the overlapping bounding boxes are localized
`and managed in order to delete false positives.
`The algorithm analyzes images on a frame by frame
`basis, without any temporal correlation.
`
`I. INTRODUCTION
`Vehicle detection is a central problem in intelligent
`vehicles. A lot of research teams work on this problem; most
`of the solutions use a stereo system in order to reconstruct
`a 3D approximation of the framed scene [1] [2] [3] [4] [5]
`[6]. A 3D reconstruction is also possible with a monocular
`system; in fact the use of two subsequent frames can create
`a sort of stereo system, but this process needs a very precise
`pitch and yaw measurement [7]. All these methods provide
`a good precision in vehicle distance detection, incidentally
`the use of a single camera together with a precise calibration
`provides a satisfactory precision.
`Some methods use models to identify a vehicle; many dif-
`ferent models have been used, ranging from trivial models
`to complex ones, like deformable models that add details of
`the car approaching the camera [8], or 3D models that take
`into consideration vehicle misalignment with the camera.
`[9]. All these methods need models that match different
`vehicle types.
`
`The authors gratefully acknowledge the grant provided by ATA as a
`support to this research.
`
`The search for vehicle features provides a simplified
`way of localizing vehicles. For example symmetry is a
`characteristic that is common to most vehicles. Some groups
`have already used symmetry to localize vehicle [11] [12],
`they have tried out a lot of methods to find symmetry on
`images: using edges, pixel intensity, and other features. It
`has been chosen to use symmetry to search for vehicles,
`improving a previous version of an algorithm described in
`[10]. The limit of the old method is to be based on detection
`of a single vehicle, only in predefined area of the image
`in front of the camera. The goal is to enlarge the search
`area and to detect all vehicles present in the image. The
`algorithm is as simple as fast nonetheless it reaches good
`results.
`In the next section the problem and the vehicle setup are
`explained, then in section III a description of the approach
`and the algorithm that has been developed can be found.
`In section IV the results are discussed and some final
`considerations exposed.
`
`II. PROBLEM AND SETUP
`
`The goal is to detect all vehicles in the scene: large,
`small, far and close ones. Unfortunately a search for ve-
`hicles with variable size and distance can be very time
`consuming. Therefore a specific algorithm together with
`efficient optimization have been designed. Furthermore
`since the computation of vehicle distance is also required
`a precise camera calibration is mandatory. Indeed drifts
`from the static calibration are often experienced, due to
`vehicle movements; vehicle pitch adds noise to the distance
`measurement.
`In the current setup the camera is mounted inside the
`cabin near the rear-view mirror; the camera aperture is 45
`degree, the image resolution is 640 x 480 while the working
`part of the image is 640 x 300 pixels. The vehicle is shown
`in figure 1.
`The test sequences were acquired on exurban roads and
`highways (fig. 2).
`
`III. ALGORITHM DESCRIPTION
`
`The algorithm is divided into the following parts:
`
`VALEO EX. 1006
`
`

`

`Fig. 1. The test vehicle.
`
`Interest areas for the detection of: far (left), medium distances
`Fig. 3.
`(center), close (right) vehicles.
`
`(a)
`
`(b)
`
`Fig. 2.
`
`Input images: (a) exurban road (b) highway.
`
`A. creation of three areas of interest; on each area the
`following steps are performed:
`– vertical symmetry computation,
`– interesting columns identification,
`– bounding boxes generation and
`– position and size filtering.
`B. Result mixing and real distance and size computation.
`
`A. Areas of interest
`Computing the symmetry on the whole image is very
`time consuming; moreover close vehicles can present too
`much details posing a significative problem in symmetry
`computation. Besides this, in order to speed up execution
`the algorithm considers three different interest areas, the
`first one for the detection of far vehicles (40-70 mt.), the
`second one for vehicles at medium distances (25-50 mt.)
`and the last one for close vehicles (10-30 mt.) (also see
`[13] for distance-based subsampling). The first area is a
`cropped version of the original image; the other areas are
`obtained subsampling the original image, or a part of it, at
`different steps in order to partially remove details, all areas
`are reduced to a fixed size (200 x 100 pixel). Figure 3 shows
`the three interest areas.
`The following parts of the algorithm are applied to the
`three images.
`
`B. Symmetry computation
`Symmetry calculation is the main part of the algorithm,
`and the most
`time consuming as well. Only binarized
`edges are used in order to reduce execution time: grey
`levels symmetry is very time consuming and does not
`provide more information than edges symmetry. First of
`all the Sobel operator is used to find edges module and
`
`orientation, then three images are built, one with binarized
`Sobel modules, one with all the almost-vertical edges and
`the last with all the almost-horizontal edges. The idea is to
`find all the vertical and horizontal edges, even the weakest.
`Therefore a very low threshold is used on edges modules;
`vertical edges are labeled considering their orientation.
`The symmetry is computed for every column of the three
`images, on different sized bounding boxes whose height
`matches the image height and with a variable width ranging
`from 1 to a predetermined maximum value. The computed
`value is saved in a 2D data structure (hereinafter referred to
`as an image) whose coordinates are determined as follows:
`the column is the same as the symmetry axis and the row
`depends on the considered bounding box width (fig. 4).
`
`(a)
`
`(b)
`
`Fig. 4. Symmetry calculation: (a) for a single column width and (b) for
`the largest width.
`
`Symmetry is calculated as
`simm = s2
`n
`where s is the number of the symmetric points, and
`n is the number of all white pixels;
`this operation is
`used on the three images of the edges (binarized Sobel,
`vertical edges, horizontal edges); two vertical edges are
`considered symmetric if their orientation is opposite. Since
`this operation has to be executed for a large number of times
`(about 400000 times for every area) it must be as simple and
`quick as possible. A new image is created with a pixelwise
`AND of the horizontal and vertical edges symmetry images.
`This image is used to search for interesting columns.
`
`

`

`C. Interesting columns
`An interesting column is defined as having a high sym-
`metry in: (i) the image that contains the result of Sobel
`binarization or in, (ii) the image that contains the AND
`between symmetry of horizontal and vertical edges. A
`columnwise histogram is then used to locate candidate
`columns. In correspondence to these columns the vertical
`edges symmetry is checked to obtain the expected vehicle
`width. More specifically if a high value of symmetry is
`present for small widths too, it means that the algorithm has
`detected a small object; in this case the column is discarded.
`Figure 5 shows an example: the leftmost peak is discarded
`because it presents a high symmetry value also for small
`widths, on the other hand the rightmost peak presents an
`appreciable symmetry value only for widths above a certain
`size.
`
`Vertical edges symmetry: the horizontal axis represents the
`Fig. 5.
`position of symmetry axis while the vertical axis represents the width
`of the symmetry box (small on the top and large on the bottom of the
`image).
`
`The Sobel binarized symmetry of the discarded columns
`are further checked: if the histogram is greater than a very
`high threshold the column is considered valid, otherwise it is
`definitely discarded; this step is used to locate vehicles that
`do not present perfect vertical edges, for example vehicles
`misaligned with the camera (fig. 6).
`
`Fig. 6. Columns valuation: only red and yellow columns are processed.
`
`D. Bounding Boxes generation
`Up to now the algorithm gives information about the
`vehicle’s center position only but since vehicle width is
`needed as well, a precise bounding box detection is manda-
`tory. For each peak in the vertical edges symmetry image
`that survived the previous filterings the width of the box is
`given by the distance between the peak and the top of the
`symmetry image; the box is centered in the column (fig. 7).
`Otherwise -in case this is not possible- a columnwise
`histogram of the number of edges is considered to detect
`the box width (fig. 8).
`
`Fig. 7. Box width computed from peak height.
`
`Fig. 8. Histogram of edges and relative box width.
`
`The shadow under the car is searched for in order to
`find the box base. It is defined as a horizontal edge, but
`since other shadows, like bridges’ ones, could be present
`on the road as well, and the algorithm looks for a high
`concentration of edges above the horizontal edge; if no base
`can be detected the column is discarded. The search for
`vehicle roof is not performed and a rectangle with aspect
`ratio equal to 4
`3 is displayed.
`
`E. Filters
`
`Unfortunately not all the detected boxes are correct: some
`false positives caused by road signs or other objects in the
`scene can be present as well. Two filters are used in order
`to discard some false positives; the former analyzes the
`box position and deletes the boxes too far from the camera
`taking into account perspective constraints, namely with the
`base too high in the image. The other one removes boxes
`too large or too small that probably do not identify a vehicle
`(fig. 9).
`It is also possible that a car is detected in more than one
`column, so overlapping boxes may be present. Only one
`box per vehicle is expected, so a further step is required
`to merge similar boxes and eliminate redundant ones: the
`largest box is preferred. Furthermore when two rectangles
`with similar size have their base at the same height an
`average is computed, and this new box is considered.
`
`

`

`are clear on both sides the algorithm works fine and detects
`the vehicles properly (fig. 12.a), otherwise when other edges
`are present, vehicles width may be incorrect (fig. 12.b).
`
`(a)
`
`(b)
`
`Filters: (a) position and (b) size. Bottom row: original images,
`Fig. 9.
`top row:enlargement of details showing the filtered boxes.
`
`F. Results mixing
`Three different lists of vehicles are produced, one for
`far vehicles, one for medium distance vehicles, and one for
`close vehicles. All the identified vehicles must be fused in
`a single list that contains all the vehicles in the image. The
`three lists are checked and the overlapping boxes removed:
`filters based on size (similar to the one described in III-E)
`are applied to bounding boxes coming from different lists.
`Distance and size of vehicles are computed thanks to
`image calibration; all the bounding boxes with unreal size
`are finally eliminated.
`
`IV. DISCUSSION OF RESULTS AND CONCLUSIONS
`The algorithm detects most of the vehicles in the provided
`test sequences, with a reduced number of false positives.
`Vehicles that proceed just in front of the camera are detected
`with a good precision (fig. 10.a), and also oblique ones
`can be identified (fig. 10.b). Trucks are detected as well
`since their square shape matches the algorithm assumptions
`(fig. 10.c). Also vehicles coming on the opposite lane can
`be detected (fig. 10.d).
`Unfortunately sometimes the vehicle size can not be
`identified with a sufficiently high precision because of
`the simple histogram method. This basically happens with
`misaligned vehicles especially (fig. 11.c). Even if filters to
`remove false positives have been introduced, there are cases
`in which ghost vehicles are detected. Figure 11.a shows
`an example of false detection induced by the presence of
`vertical edges above a horizontal shadow line. Sometimes
`the base of the vehicle is detected with an error because of
`the presence of a black bumper (fig. 11.b).
`Bridges shadows are very difficult to manage because
`they can generate a lot of false positive and can also
`interfere in base detection (fig. 11.d). Figure 11.d also shows
`a good performance of the algorithm in presence of a simple
`overlapping of two cars: both cars are detected.
`The problem of misaligned vehicles (like overtaking
`vehicles or vehicles coming in the opposite direction) is the
`most critical aspect of this method. The algorithm does not
`always detect oblique vehicles, and when it does, vehicle
`width may not be precise. When vehicles rear vertical edges
`
`(a)
`
`(b)
`
`Fig. 12. Misaligned vehicles: (a) correct detection, (b) width error.
`
`The algorithm works efficiently on simple images. Prob-
`lems arise when a large number of vehicles overlap or in
`urban scenes, where road infrastructures, road signs and
`shadows make the scene too complex.
`Execution time is 120 ms with an Athlon XP 1.8 GHz;
`since 75% of the time is used for symmetry computation,
`the next research step will be the optimization of this part,
`moreover, a quantitative performance measurement is being
`carried out for comparison with other approaches.
`REFERENCES
`[1] U. Franke, “Real-Time Stereo Vision for Urban Traffic Scene Un-
`derstanding,” in Procs. IEEE Intelligent Vehicles Symposium 2000,
`(Detroit, USA), pp. 273–278, Oct. 2000.
`[2] H. Li, S. Bouaziz, and F. Devos, “An Embedded System for Au-
`tonomous Collision Avoidance and Line Tracking Using Artificial
`CMOS Retina Sensor,” in Procs. IEEE Intelligent Vehicles Sympo-
`sium 2000, (Detroit, USA), pp. 273–278, Oct. 2000.
`[3] S. Heinrich, “Fast Obstacle Detection using Flow/Depth Constraint,”
`in Procs. IEEE Intelligent Vehicles Symposium 2002, (Paris, France),
`June 2002.
`[4] R. Labayarde, D. Aubert, and J. P. Tarel, “Real Time Obstacle
`Detection in Stereovision on non Flat Road Geometry through
`’V-disparity’ Representation,” in Procs. IEEE Intelligent Vehicles
`Symposium 2002, (Paris, France), June 2002.
`[5] C. Kn¨oppel, A. Schanz, and B. Michaelis, “Robust Vehicle Detection
`at Large Distance Using Low Resolution Cameras,” in Procs. IEEE
`Intelligent Vehicles Symposium 2000, (Detroit, USA), pp. 267–272,
`Oct. 2000.
`[6] A. Bensrhair, M. Bertozzi, A. Broggi, A. Fascioli, S. Mousset,
`and G. Toulminet, “Stereo Vision-based Feature Extraction,” in
`Procs. IEEE Intelligent Vehicles Symposium 2002, vol. 2, (Paris,
`France), pp. 465–470, June 2002.
`[7] Z. Hu and K. Uchimura, “Tracking Cycle: A New Concept for
`Simultaneously Tracking of Multiple Moving Objects in a Typical
`Traffic Scene,” in Procs. IEEE Intelligent Vehicles Symposium 2000,
`(Detroit, USA), pp. 233–239, Oct. 2000.
`[8] S. Denasi and G. Quaglia, “Obstacle Detection Using a Deformable
`Model of Vehicles,” in Procs. IEEE Intelligent Vehicles Symposium
`2001, (Tokyo, Japan), pp. i145–150, May 2001.
`[9] K. Fleischer, H. H. Nagel, and T. M. Rath, “3D-Model-based-Vision
`for Innercity Driving Scenes,” in Procs. IEEE Intelligent Vehicles
`Symposium 2002, (Paris, France), June 2002.
`[10] M. Bertozzi, A. Broggi, A. Fascioli, and S. Nichele, “Stereo Vision-
`based Vehicle Detection,” in Procs. IEEE Intelligent Vehicles Sym-
`posium 2000, (Detroit, USA), pp. 39–44, Oct. 2000.
`[11] A. Kuehnle, “Symmetry-based vehicle location for AHS,” in
`Procs. SPIE - Transportation Sensors and Controls: Collision Avoid-
`ance, Traffic Management, and ITS, vol. 2902, (Orlando, USA),
`pp. 19–27, Nov. 1998.
`
`

`

`(a)
`
`(c)
`
`(b)
`
`(d)
`
`Fig. 10. Results: (a) a car driving in front of the camera (b) a misaligned car (c) a truck (d) vehicles on the opposite lane.
`
`(a)
`
`(c)
`
`(b)
`
`(d)
`
`Fig. 11. Errors: (a) false positive (b) wrong base position (c) wrong size (d) another wrong base position.
`
`[12] T. Zielke, M. Brauckmann, and W. von Seelen, “Intensity and Edge-
`based Symmetry Detection with an Application to Car-Following,”
`CVGIP: Image Understanding, vol. 58, pp. 177–190, 1993.
`
`[13] V. Graefe and W. Efenberger, “A Novel Approach for the Detection
`of Vehicles on Freeways by Real-time Vision,” in Procs. IEEE
`
`Intelligent Vehicles Symposium‘96, (Tokyo, Japan), pp. 363–368,
`Sept. 1996.
`
`

`

`QIEEE
`
`Advancing Technology
`for Humanity
`
`DECLARATION OF GERARD P. GRENIER
`
`1, Gerard P. Grenier, do hereby state that:
`
`l.
`
`I am currently Senior Director of Publishing Technologies for the Institute of Electrical
`and Electronics Engineers (IEEE), 445 Hoes Lane, Piscataway, New Jersey.
`
`2.
`
`I have been asked to confirm certain dates regarding the following article (the “Article”):
`
`A. Broggi, et al., “Multi-Resolution Vehicle Detection using Artificial Vision,” 2004
`IEEE Intelligent Vehicles Symposium, June 14-17, 2004.
`
`3. A true and correct copy of the Article accompany this declaration as Exhibit A.
`
`4.
`
`IEEE is a neutral third party to the dispute in this IPR.
`
`5. Neither I nor IEEE itself is being compensated for this declaration.
`
`6. The Article was published by IEEE. Copies of the conference proceedings were made
`available to attendees of the conference. The Article is currently available for public
`download from the IEEE digital library, IEEE Xplore (www.ieeexplore.ieee.org).
`
`7.
`
`IEEE maintains records of the dates on which articles were first made available for public
`download. These records are maintained as part of the ordinary course of business of the
`IEEE, are updated regularly as articles are published. I have personal knowledge of the
`records.
`
`8.
`
`I have reviewed the IEEE records relating to the Article.
`
`9.
`
`IEEE’s records confirm the following:
`
`a) “Multi-Resolution Vehicle Detection using Artificial Vision” was presented at the
`2004 IEEE Intelligent Vehicles Symposium which occurred June 14-17, 2004.
`
`b) IEEE registered the copyright to this conference on January 4, 2005.
`
`I declare under penalty of perjury that the foregoing statements are true
`
`Executedon:
`
`_93.21.44.141£151-
`
`(I curr
`
`‘1.
`
`‘_ Z X. if}
`
`
`
`“*4
`
`445 Hoes Lane Piscataway, NJ 08854
`
`

`

`EXHIBIT A
`
`

`

`6/30/2014
`
`IEEE.org
`
`iEEE Xplore Abstrath Multi=resolution vehicle detection using artificial vision
`
`IEEE Standards
`
`IEEE Spectrum
`
`More Sites
`
`Cart (0) Create Account Sign In
`
`IEEEXp/ore“w
`
`mogul. unmtlav
`
`A
`
`F
`
`-1
`
`t't t'
`
`(ilk-SEC:
`
`"'.::..'..:;:::.;...;:
`
`Athens/Shibboleth
`
`WEEE
`
`Back to Results
`
`|
`
`B'owse Conference Publications > Intelligent Vehicles Symposiu .
`
`Mum-resolution .vehlcle detectlon
`usmg artIfICIal Vlsmn
`
`a Full Text
`alga." or purchase
`
`Broggi. A.
`
`'. Dipl. di lngegnen‘a dell"lnf., Parma Univ, ltaly ; Carri, P. ; Antonella, P.C.
`
`Abstract
`
`Authors
`
`References
`
`Cited By
`
`Keywords
`
`Metrics
`
`Similar
`
`This paper describes a vehicle detection system using a single camera. it is based on the search for
`areas with a high vertical symmetry in mulii-resolution images; symmetry is computed using different
`sized boxes centered on all the columns of the interest areas. All the columns with high symmetry are
`analyzed to get the width of detected objects. Horizontal edges are examined to find the base of the
`vehicle in the individuated area. The aim is to find horizontal lines located below an area with sufficient
`
`amount of edges. The algorithm deletes all the bounding boxes which are too large, too small, or too far
`from the camera in order to decrease the number of false positives. All the results found in different
`interest areas are mixed together and the overlapping bounding boxes are localized and managed in
`order to delete false positives. The algorithm analyzes images on a frame by frame basis, without any
`temporal correlation.
`
`Published in:
`
`intelligent Vehicles Symposium, 2004 IEEE
`
`0
`
`Tweet
`
`0
`Share
`
`Date of Conference:
`1447 June 2004
`
`Page(s):
`310 - 314
`
`Print ISBN:
`0-7803-8310—9
`
`INSPEC Accession Number:
`8160826
`
`Digital Object Identifier :
`10.1109/IVS.2004.1336400
`
`Publisher:
`IEEE
`
`Change Username/Password
`Update Address
`
`Payment Options
`Order History
`
`Communications Preferences
`Profession and Education
`
`us 8- Canada: +1 800 678 4333
`Worldwide: +1 732 981 0060
`
`Access Purchased Documents
`
`Technical Interests
`
`Contact & Support
`
`About IEEE Xplore Contact Help
`
`Terms of Use Nondiscrimination Policy
`
`Site Map
`
`Privacy e. Opting out of Cookies
`
`Sign In | Create Account
`
`http://ieeexplore.ieee.org/ml/articleDetails.jsp?tp=&arnumber=1336400&q ueryTe)t%3DMulti-Resolution+Vehicle+Detection+using +Artiiicial+VIsion
`
`1/1
`
`