semi-automatic road tracking using parallel angular

Intelligent Automation and Soft Computing, Vol. 18, No. 8, pp. 1009-1021, 2012 Copyright © 2012, TSI® Press Printed in the USA. All rights reserved

SEMI-AUTOMATIC ROAD TRACKING USING PARALLEL ANGULAR TEXTURE SIGNATURE XIANGGUO LIN1*, JING SHEN2, YONG LIANG3

1

Key Laboratory of Mapping from Space of State Bureau of Surveying and Mapping Chinese Academy of Surveying and Mapping 28, Lianhuachixi Road Haidian District Beijing, P.R. China 100830 2

Research Center of Government Geographic Information System Chinese Academy of Surveying and Mapping Beijing100830, P.R. China 3

School of Information Science and Engineering Shandong Agricultural University 61, Dazong Road, Tai’an Shandong Province P.R. China 271018

* Corresponding author, Tel: +86-10-63880576, Fax: +86-10-83956425 Email: [email protected]

ABSTRACT—Road tracking is a promising technique to increase the efficiency of road mapping. In this paper, a semi-automatic road tracker, Parallel Angular Texture Signature (PATS), is presented. The tracker is object-oriented in some sense, because it makes best use of the texture signature of road primitives on high-resolution remotely sensed imagery. Our tracker uses parabolas to model the road trajectory and predicts the position of next road centerline point. It employs PATS to get the moving direction of current road centerline point, and it will move on one predefined step along the moving direction to reach a new position, and then it uses curvature change to verify the newly added road point. Moreover, we also build compactness of Angular Texture Signature polygon to check whether the PATS is suitable for subsequent tracking. Repeat the above steps until the whole task is finished. Extensive experiments demonstrate that the proposed tracker is capable of efficiently extracting most of main roads from medium and low resolution imagery, and reliably and robustly extracting most of ribbon roads from high resolution SAR and optical imagery. Key Words: semi-automatic; road tracking; profile matching; template matching; angular texture signature

1009

1010

Intelligent Automation and Soft Computing

1. INTRODUCTION Extraction of the roads from digital aerial/satellite imagery is not only scientifically challenging but also of major practical importance for spatial data acquisition and update for geographic information system (GIS) databases, navigation, transportation plan and management, estimation of air pollution etc. Traditionally, manual plotting is deployed in road extraction, but it is time consuming and expensive, so automatic or semi-automatic acquisition and update of road data is greatly needed, especially after the availability of high resolution satellite imagery such as IKONOS and QuickBird. In the last three decades, there was extensive research on automatic or semi-automatic road extraction from aerial/satellite imagery including optical imagery, Synthetic Aperture Radar (SAR) data, LIDAR data, and image sequence taken from mobile vehicles. As a result, many strategies, methodologies and algorithms for road extraction were presented and they reached various extents of success [1]. The main problem of fully automatic approach is that it needs strict hypothesis of road characteristics. However, road characteristics vary considerably with ground resolutions, road types, and density of surrounding objects, light conditions etc. Therefore, the quality of automatic extraction is not sufficient due to some missed or erroneous segments and much necessary post-edition. On the contrary, semi-automatic approach that retains the human operator in the loop where computer are used to assist human performing is considered to be a good compromise between the fast computing speed of a computer and the efficient interpretation skills of an operator. Correspondingly, quite a lot of promising approaches for semiautomatic road extraction have been proposed so far. In [2], profile correlation is utilized to track roads on high resolution aerial imagery. In [3], edge detection is used to extract ridges. In [4] edgebased and profile correlation based approaches are combined to follow a road. In [5], least-squares profile matching is employed to track roads. In [6], least-squares template matching is used. In [7], road seeds are connected by dynamic programming. In [8], snakes is used to optimize the path of road seed points. In [9], feed forward neutral network is used to optimize the path of seed points. In [10], minimum cost is employed to follow a path. From the survey above, we can see that most of the above semi-automatic methods focus on specific subtasks and situations such as specific image resolution, image type, road type, and the complexity of the image scene. Furthermore, there is still no one single method that will be successful in extraction of all types of roads from any image in any scene. In this paper, a robust semi-automatic road tracker is proposed to extract the roads with many occlusions of vehicles and shadows of trees based on the road characteristics in remotely sensed imagery. Our tracker can make best use of the texture signature of road primitives to overcome the side-effects of shadows and vehicle occlusions on roads, and it employs parabola to fit the road path and predict the position of next road centerline point one step ahead, and it also utilize the curvature change of adjacent two road points to verify the new added road centerline point.

2. ROAD MODEL AND PRINCIPLE OF OUR DETECTOR 2.1. Road Characteristics

The road characteristics consist of geometrical, photometric (or reflective), topological, functional and contextual (or semantic) characteristics [5, 7, 8]. Reviewing the existing works on road extraction, we find that the given road characteristics play a very important role in extraction, but the road characteristics greatly depend on the sensor type and spatial resolution. In this paper, we focus on the extraction of ribbon roads from high-resolution optical and SAR imagery.

Semi-Automatic Road Tracking Using Parallel Angular Texture Signature

1011

2.2 Principles of Angular Texture Signature

Varirance

A texture measure described in [11], is redefined in [12, 13], as shown in Figure 1. At each pixel p of a panchromatic image, T(α, ω, p) is defined as the variance from the mean for a rectangular set of pixel of width ω around the point p whose principal axis lies at an angle of α from horizontal. This measure is computed for a set of angles α0, …, αn. Figure 1 (a) shows the rotating templates for a single point. In order to well illustrate the meaning, we select the point p at the center of a crossing. At the point p, the angular texture signature (ATS) is defined as the set of values { T(α0, ω, p), T(α1, ω, p), T(αn, ω, p)}. The graph of an ATS for the above single point p is shown in Figure 1(c). The local minima on this graph correspond to the most likely directions of the road at p. At each pixel p, the number k and location of the strong local limits are computed from the ATS. For example, the signature shown in Figure 1(a) has 4 significant limits. We refer to the number k of limit as the degree of the pixel. The texture measures that are used in road detection are: the degree of the pixel and the direction of the limit. In this sense, ATS is an objectoriented tracker by using the property of primitives of a road. If a road segment has a good contrast with adjacent area, the mean of the rotating template can take the same effect as a discriminator to indicate the potential road direction, as shown in Figure 1(d). 7000 6000 5000 4000 3000 2000 1000 1

4

7

10 13 16 19 22 25 28 31 34 Anglular Index

(a)

(c)

Mean Value

200 180 160 140 120 100 1

4

7

10 13

16 19 22 25 28 31

34

Angular Index

(b)

(d)

Figure 1. Angular Texture Signature. (a) Texture is computed over a set of rectangular template rotating around a road centerline point with equal interval (Note that the width of the template is as wide as the road’s width while the length of the template is as long as four times of road’s width, and there are 36 templates in total but only odd ones are displayed), (b) The value of each template and the ATS polygon, (c) The graph of the ATS taking variance as a measure, (d) The graph of ATS taking mean as a measure.

1012

Intelligent Automation and Soft Computing

ATS is used to automatically extract urban road networks in [12] and [13], and the results suggest its potential applications on road tracking. Its merits thanks to that it makes best use of the spectral characteristics and textural characteristics of roads on high resolution images. However, it is hard to populate in complex scenes, but modified ATS has extensive applications. In this paper, we extend and redefine ATS to fit to semi-automatic road tracking. In semi-automatic road tracking, initial road direction is given, adding that the curvature change has a maximum, so the rotating angle range is not necessary to be so large. The ATS is redefined as follows. At each road centerline point p of a grey image, T(α, ω, h, p) is defined as the mean, standard deviation, variance or entropy for a rectangular set of pixels of width ω and height h around pixel p whose principal axis lies at an angle of α from the potential road direction θ . Note that standard deviation, variance or entropy takes the same effects. This measure is computed for a set of angles α0, α1,…,αn at pixel p. Angles α0, α1,…,αn symmetrically locate two sides of angleθ with equal interval σ and having the maximum difference ξ with θ . At the point p, the ATS is regarded as the set of values { T(α0, ω, h, p) , T(α1, ω, h, p)…,, T(αn, ω, h, p)}. Figure 2(a) displays an ATS with σ =10° and ξ =40°, and Figure 2(b) shows the corresponding ATS values (mean as a measure) related to Figure 2(a). The direction of the limit of ATS is regarded as the road direction. If the ATS takes the variance or entropy as measure, the direction of local minimum is taken; while if the ATS takes the mean as measure, the direction of local maximum is taken for bright roads and the direction of local minimum is taken for darker road. Once the road direction is given, move on one step along the direction and iterate the above process until the tracker fails or reaches to another tracked road trajectory or reaches to the boundary of the image.

Mean Value

100 80 60 40 20 0

1

2

3

4

5

6

7

Angular Index

(a)

(b)

Figure 2. Angular Texture Signature. (a) ATS used to track road, (b) The graph of ATS taking mean as a measure.

After conducting a lot of experiments, we find that the along-track sudden change of radiometric or reflective property of road surface doesn’t terminate the tracking process and affect the accuracy of tracked result. We also verify that it will not affect the tracked trajectory if the shadow looks like the appearance of radiometric change. At the same time, ATS can be deployed to track linear roads with brighter intensity in low and medium resolution imagery. But in reality, the road conditions are more complex especially containing the appearance of shadows of colonnades and buildings, occlusion of vehicles, and change of building material. Once emergence of such disturbance, the above ATS tracker fails so frequently that it is no longer fit for following a road. But when we make a close look at the high resolution optical imagery, we can conclude that most of shadows, conclusions occur only on half part of the road. And we also verify that the ATS of half part of the road looks the same as the ATS of the whole road segment if the

Semi-Automatic Road Tracking Using Parallel Angular Texture Signature

1013

road condition is ideal. That provides a clue on how to track road when shadow and occlusion occurs. Then we resort to employ two cooperative ATSs of right and left half part of a road segment to track road. When one ATS fails, the other may still work, as shown in Figure 3. We name the extended ATSs as Parallel ATSs (PATS).

Figure 3. The graph of P ATS

2.3 Road Trajectory Model

Road trajectory model is helpful to predict the most possible position of road point one step ahead and to guide the tracking go through bad road conditions. Road trajectory can be modeled by a B-splines [8], Kalman filter process [5], and a parabola [4]. We resort to fit a parabola to the most latest tracked twenty road points. The parabola in the x⊥ y plane can have arbitrary orientation, having an equation of the form:

Ax 2 + Bxy + Cy 2 + Dx + Ey + F = 0

(1)

where x and y are coordinates of points on the parabola, A, B, C, D, E, F are parameters. We don’t use this equation since the parametric form is more convenient for our purpose. We represent the road path parametrically as two separate functions x(l) and y(l) where l is the total length in steps that we have traversed along the road’s path. We use multiple linear regressions with l2 and l as the independent variables to fit parabolas to x(l) and y(l), getting approximate functions:

X (l ) = a1l 2 + b1l + c1 Y (l ) = a 2 l 2 + b2 l + c 2

(2)

where x(l) and y(l) are the x coordinates and y coordinates of road centerline points, a1, b1, c1, a2, b2, c2 are the parameters, and l=i*ω

(3)

where i is the number of a road centerline point in the queue of road centerline points, and ω is the step length, which means the step length is preset to road width ω in this paper. To get the most possible potential direction of the road, we resort to compute x(l+1) and y(l+1), which are the most possible location of the next road centerline point. Given that the road segment has a maximum curvature, then the change of the curvature of two adjacent road points must be less than a predefined threshold, T. The curvature K at some point on the parabola can be computed as follows:

1014

Intelligent Automation and Soft Computing

K=

2 | b1c 2 − c1b2 | 2

2

[4(c1 + c 2 ) 2 l 2 + 4(c1b1 + c 2 b2 )l + b1 + b2 ]3 / 2

(4)

where a1, b1, c1, a2, b2, c2 are introduced in Eq. (2).

2.4 Compactness of ATS Polygon

When we take a closer look at the ATS rotating a full 360°, we can find some interesting links between the shape of the ATS polygon and corresponding pixel types. To form the ATS polygon, instead of plotting the ATS values for each direction along a horizontal line, we plot the ATS values around the pixel under consideration with corresponding direction and link the last point to the first point [13]. The resulting polygon is called the ATS polygon. Figure 1(b) shows the calculated ATS for some a interesting pixel with the ATS polygons in red colour. If the road has a good contrast with its surrounding objects, the polygon usually looks like an ellipse or ∞ -shape. The compactness of ATS is defined as the compactness of the ATS polygon using Eq. (5). The compactness tells us whether the ATS polygon looks like a circle. A circle-like ATS polygon usually means that the tracker is no longer fit for following the road, and manual plotting is needed instead.

ATScompactness =

π⋅A P2

(5)

where A and P are the area and perimeter of the ATS polygon, respectively. Note that our program will calculate the compactness of ATS at regular intervals to verify whether the ATS or PATS is still reasonable to track a road.

3. THE OVERALL STRATEGY OF PATS TRACKER Semi-automatic road extraction here is undertaken as follows: (1) Preprocess the input image If the original image doesn’t have a good contrast between road and other features, it needs stretching. Then the image is convolved with a Gaussian filter to smooth the image and reduce the high-frequency noise.

x2 + y2 G = exp(− ) 2σ 2

(6)

where σ is a predefined parameter based on image resolution. (2) Operator detect a road segment A human operator has to identify a short part of a road axis, and this road part serves as initialization for an automatic tracker. The tracker needs a starting point on the road centerline and a second point to define the direction of the road and a third point to define the width ω of the road. Then the starting point and the direction of the road are stored in a list. Predict the next road position and then calculate the most possible potential road direction. In our scheme, user can initialize the road in another approach inputting a queue of points on each roadside if there is no salient centerline. The program will get the initial starting point, road direction and width by parallelogram fitting. (3) Compute PATS

Semi-Automatic Road Tracking Using Parallel Angular Texture Signature

1015

From the last road centerline point in the trajectory list, the program generates two new points higher or lower than the last point along the perpendicular of the potential road direction with a distance ω/4 respectively. A rectangular template is formed with width ω/2 and height 2*ω respectively around the above two points. At discrete intervals about each pixel, the signature is calculated. Selecting which texture signature as the measure of ATS, it should judge by the road conditions. If road is more homogenous, taking variance, strand deviation or entropy as the measure while taking mean as the measure if the road has obvious contrary with surroundings. If the limit of ATS surpasses the preset threshold, T0, then resort to manual plotting. (4) Compute Calculate the

ATS compactness and move on one step ahead

ATS compactness of points under consideration by Eq. (5). If the compactness

values of two parallel points are both larger than predefined threshold, T1, it tells us that the ATS is not suitable to tracking the road any more, and it needs manual plotting. Otherwise the direction of the limit is regarded as the road direction, and move the road trajectory one step ahead. We must note that the step length is equal to road’s width in our system, but it is also can be adjusted (5) Compute the curvature change Calculate the curvature of the new added seed point by Eq. (4), and compare it with the curvature of last point, if the difference is larger than predefined threshold T2, delete the new point from the road trajectory list, and then resort to manually plot. Otherwise, predict the next road position by parabola equation and iterate from step (3) if the trajectory doesn’t reach to another trajectory or the boundary of the image. Once the user accomplishes one road segment or the tracker reaches to one tracked road or the boundary of image, initializes another road segment and restart from step (2) again until all roads are tracked. From the operator point of view, the procedure is as follows: the operator has to initialize the tracker by three input points to indicate the starting, the moving direction and the width of the road segment, and then the tracker is triggered. Whenever the internal evaluation of the tracking tool indicates that the tracker might lose the road axis or be no longer fit, it demands for interaction of the operator. Then the operator has to confirm the tracker or must edit the extracted wrong road parts and put the tracker back on the road.

4. EXPERIMENTS AND EVALUATION We develop a prototype system on a computer with Pentium 4 2.00 GHz CPU and 1.00 GB RAM using VC++ 6.0 over Windows XP operation system, and the ATS, PATS and other algorithms such as profile matching, template matching, snakes are implemented. Various images in urban, suburban, or rural area and images with different resolutions are tested. We declare that our algorithm only use the information of one grey image, and only red bands are deployed for colorful experiment images, all testing data are made available from Chinese Academy of Surveying and Mapping.

4.1 Extracting Ribbon Roads from SAR Image by PAT

Property of roads on high resolution SAR imagery is not so disturbed by vehicle occlusions, but spectacles exist due to the imaging mechanism of SAR, which limits the ATS taking variance as a measure, but mean value is still effective when the most of the road surfaces on SAR imagery are darker than surrounding features. This experiment uses a airborne SAR image Of He’fei City, China, with size 3955X3859 pixels, and its spatial resolution is 0.3m. There are two salient main roads on the image, a brighter one and a darker one, and they both have a good contrast with adjacent area. The tracking takes

1016

Intelligent Automation and Soft Computing

about 321 seconds, and there are 5 errors in the process. Once an error happens, manual intervention is necessary to delete the wrongly tracked points and to guide the tracker back to the road centerline again. From the result, we can see that the accuracy of tracked trajectories is guaranteed despite with some surge at some local area. And the errors are encountered where there is a turning or the road width suddenly changes, but we must note that there is no stop at crossing of the two roads. In this experiment, we meet the situation that compactness of ATS functions when an error occurs but the operator can’t react as soon as possible.

4.2 Extracting Roads from Optical Imagery 4.2.1 Extracting salient road centerline and lateral edges by PATS Our tracker is designed to track road in complex scenes, but it also works as well as other current tracker for roads with good conditions. Based on the principle of PATS, it is also useful to track edges of roads. The test is made on a Quickbird panchromatic image of Huai’rou County in Beijing, China, with size 355X1066 pixels. On this image, there is a darker centerline on the homogenous brighter road surface with a darker background. The tracking takes about 150 s, and there is no stop for each line. From the result, we can see that the extracted centerline is deviating its real position at the turning part. We use least square template matching to track the centerline of the same road with same moving step length, and the result is similar to the one of PATS. It suggests that it is the rectangular template that causes the above problem rather than parabola. But if we shorten the moving step further, the problem is partially solved. If the moving step is so short, the tracking takes more time. So there must be a trade-off between accuracy and consuming time. In practical application, if road centerline is hard to extract but easy for edges, then edges can be detected by PATS instead of centerline. 4.2.2 Extracting road centerlines of highways by PATS Highways are maintained with some curvature and are usually disturbed by some vehicles, shadows, and change of building material. Therefore it is not so easy to track highways. A 0.2m resolution true color airborne image is used with size of 2543 by 33346 pixels acquired by DMC over Tianjin area. The spectral reflectance is not so uniform due to different building material, and there are much occlusion vehicles and shadows from the trees on the roadsides. The following goes on after initialization with one error occurrence for each road, the tracking takes about 456 s, and there is some offset from the centerline where vehicles or shadows appear, as shown in Figure10, but the deviation is acceptable. From Figure 4(c) and (d), we can see that PATS is effective and reliable to extract highways even in complex scenes. 4.2.3 Extracting a main road in urban area by PATS In urban area, the road may not have good contrast with adjacent area, and then we resort to PATS taking variance as a measure to track roads. The test image is shot by ADS40 with 20cm ground resolution, and its size is 678 by 1021 pixels. There are sparse vehicles on road surface, and some part of the road is shadowed by colonnade, and the centerline markers have different shape and colors. The tracking takes about 40 s and there are three stops in the process. From the We can see that the tracked trajectory is very close to the centerline marker.

Semi-Automatic Road Tracking Using Parallel Angular Texture Signature

(a)

(c)

1017

(b)

(d)

Figure 4． PATS is used to track highways on a DMC image. (a) Original image, (b) extracted roads, (c) local image overlapped by detected trajectory, (d) local image overlapped by detected trajectory.

4.2.4 Extracting a road in radiometrical difference between lanes and change along road by PATS Radiometric changes of road surfaces caused by material change greatly affect the effect of current road detectors, but change is very common in urban area, and even the different lanes of the same road segment have different building materials. The testing image comes from the same source as the above image, and its size is 689 by 1018 pixels. The whole procedure takes about 40 s and there are two stops. From the image, we can conclude that frequent change of building material has some side-effect on the result, but the accuracy is guaranteed thanks to the involvement of human. 4.2.5 Extracting roads with little shadow in dense building district by PATS In an orthoimage, shadows and occlusion are greatly decreased, which benefits the information extraction. The test image comes from same data source as the above one, and its size is 690 by 1016 pixels. The following takes about 226 s and there is no break of except initializations in the process. From the result, we can see that PATS is very effective in dense building area on orthoimages. Note that connections at intersections of roads are not done in our experiments.

1018

Intelligent Automation and Soft Computing

4.2.6 Extracting roads with many shadows in dense building district by PATS In very complex scene, having many shadows or occlusions, most of road trackers don’t work due to too much annoying failure and time-consuming manual participation. To testify the robustness of our tracker, an ADS40 shot image of dense building district with much shadow is used, and its size is 680 by 1012 pixels and its ground resolution is 20cm. On the image, the shadows occupy at least half the road surface. The following takes about 291 s and there are 12 stops except initializations in the process. From the result, we can see that there is large deviation between tracked trajectory and real position in some parts, but PATS do also work in this complex condition.

4.3 Experimental Evaluation

Semi-automatic systems can be evaluated by many criteria. For real-world application, it often includes user experience evaluation. Since our system is still in the simulation stage, we focused on the engineering aspect of the evaluation where the human factors components are only part of the assessment criteria. Zhou et al. [14] proposed that tracking performance was evaluated in four criteria: correctness, completeness, efficiency and accuracy. The most merits of our proposed system are that the human involvement guarantees correctness, completeness and accuracy of tracking, because the human operator supervises the whole procedure. If an error happens, the human operator can correct it immediately, deleting the wrong part and initializing a new tracking. Completeness is also guaranteed because interpretation is performing as manual plotting. Once large deviation is found in the automatic tracking, then the thread is stopped and manual plotting is undertaken to guarantee the accuracy. Our evaluation focuses on the efficiency. To evaluate efficiency, we employ saving in human inputs and saving in total time. Decreasing human inputs can relief the burden of human operator. Saving in total time can increase efficiency. Human inputs include initialization, error editing and manual tracking. Once error occurs, manual involvement consumes some time. Once the compactness efficient, curvature change or minimal variance surpasses the preset threshold, the program will pop up message box, which also wastes some time. We have recorded the times of errors, times of stops, times of mouse input and total time cost of seven experiments above. To compare the efficiency, we also manually plot the above tested images in Section 4. Manual plotting should be as accurate as possible, i. e. the mouse clicks should be on the true road axis points. Furthermore, the road should be smooth, i. e. abrupt changes in direction should be avoided and no zigzags should occur. Both of results of semiautomatic tracking and manual plotting refer to Table I. Table I suggests that the proposed tracker greatly decrease the human inputs, approximately 93% in total, but time savings significantly differ. For simple scenes such as 1 and 6, the time saving is larger than 50%. For medium complex scenes such as images of experiment 2, 3, 4 the time saving is about 40%. For complex scene such as image of experiment 7, there is no significant time saving. But if the road width is very large such as the image of experiment 3, there is no significant time saving.

Semi-Automatic Road Tracking Using Parallel Angular Texture Signature

1019

Table I. Comparison between our proposed tracker and manual plotting

Tracking errors

Human inputs

Total time costs

Human

Human 485 295 654

Semiautoma tic 42 9 12

Human

0 0 0

Semiautoma tic 1 0 0

467.4 212.4 464.5

Semiautoma tic 321.8 150.3 456.8

Experiments

Tracking stops

Human

1 2 3

0 0 0

Semiautoma tic 4 0 2

4

0

1

0

0

92

5

67.1

40.1

5 6

0 0

1 0

0 0

1 0

73 680

12 33

52.8 489.1

40.5 226.2

7

0

10

0

2

402

45

301.8

291.1

5. CONCLUSIONS The proposed tracker, PATS, is very robust, because it makes best use of the road characteristics on remotely sensed imagery. Our algorithm employs parabola to fit the road trajectory and predict the road position and moving direction and check the new added road point by the curvature change, it also utilizes compactness coefficient to evaluate the aptness of ATS polygon to go on tracking. Extensive experiments demonstrate that our proposed tracker efficiently extract linear roads in medium and low resolution imagery and reliable extract ribbon roads including highways, main roads and streets from high resolution SAR and optical imagery even in very complex scenes. But our algorithm has five obvious limitations. First, the road must be long enough. Second, the algorithm may not work when the road’s geometrical property such as width and curvature suddenly changes. Third, it can only use the cues from grey scale imagery. Fourth, it can’t track roads where vehicle is very dense in both lanes of a road. Last but not least, it needs more computing times. These limitations are currently being examined now. Further research will focus on the optimization of extracted trajectory, automatic extraction of road’s width and direction, utilization of multi-spectral information and so on.

ACKNOWLEDGEMENTS This research was funded by the Project for Young Scientist Fund sponsored by the Natural Science Foundations of China under Grant 41001280, State 863 projects under Grant G7910, the National Key Basic Research and Development Program under Grant 2006CB701303, and the China Postdoctoral Science Foundation under Grant 2010047038.

REFERENCES 1. 2. 3.

J. B. Mena. “State of the art on automatic road extraction for GIS update: a novel classification”, Pattern Recognition Letters, 2003, 24: 3037-3058. L. H. Quam. “Road tracking and anomaly detection in aerial imagery”, In: Image Understanding Workshop, London, UK, 1978, pp. 51-55. R. Nevatia and K. R. Babu. “Linear feature extraction and description”, Computer Graphics and Image Processing, 1980, 13: 257-269.

1020

4. 5. 6. 7. 8. 9.

10. 11. 12. 13. 14.

Intelligent Automation and Soft Computing

D. Mckeown and J. Denlinger. “Cooperative methods for road tracing in aerial imagery”. In: Proceedings of the IEEE Conference in Computer Vision and Pattern Recognition, Ann Arbor, MI, 1988, pp. 662–672. G. Vosselman and D. J. Knecht. “Road tracing by profile matching and Kalman filtering”. In Automatic Extraction of Man-Made Objects from Aerial and Space Images, A. Gruen, O. Kuebler and P. Agouris(Eds)(Basel: Birkhauser Verlag), 1995, pp. 265–274. T. Kim, S. R. Park, M.G Kim, S. Jeong and K.O. Kim. “Tracking road centerlines from high resolution remote sensing images by least squares correlation matching”, Photogrammetic Engineering and Remote Sensing, 2004, 70: 1417-1422. A. Gruen and H. H. Li. “Road extraction from aerial and satellite images by dynamic programming”, ISPRS Journal of Photogrammetry and Remote Sensing, 1995, 50(4): 1120. A. Gruen and H. H. Li. “Semi-automatic linear feature extraction by dynamic programming and LSB-Snakes”, Photogrammetic Engineering and Remote Sensing, 1997, 63(8): 985995. X. Y. Hu, Z. X. Zhang and J. Q. Zhang. “An approach of semi-automated road extraction form aerial images based on template matching and Neural Network”, International Achieves of Photogrammetry and Remote Sensing, Amsterdam, Netherlands, 2000, Vol. XXXIII, Part B3, pp. 994-999. V. Shukla, R. Chandrakanth and R. Ramachandran. “Semi-automatic road extraction algorithm for high resolution images using path following approach”, In: ICVGIP02, Ahmadabad, 2002, pp. 231-236. L. Gibson. “Finding road networks in IKONOS satellite imagery”, In: Proceedings of ASPRS 2003 Conference, Anchorage, Alaska, 2003, pp. 1200-1209. D. Haverkamp. “Extracting straight road structure in urban environments using IKONOS satellite imagery”, Optical Engineering, 2002, 41(9):2107-2110. Q. Zhang and I. Couloigner. “Benefit of the angular texture signature for the separation of parking lots and roads on high resolution multi-spectral imagery”, Pattern Recognition Letters, 2006, 27: 937-946. J. Zhou, W. F. Bischof and T.Caelli. “Road tracking in aerial images based on humancomputer interaction and Bayesian filtering”, ISPRS Journal of Photogrammetry & Remote Sensing, 2006, 61: 108-124.

ABOUT THE AUTHORS X. Lin received the Ph.D. degree in cartography and geographic information system Wuhan University, in 2009. He is currently a postdoctoral in Chinese Academy of Surveying and Mapping. His research interests include highresolution image processing, airborne LiDAR cloud data processing, pattern recognition, and remote-sensing applications. He coordinated 1 national research projects, and has published more than 20 national, international journals papers.

Semi-Automatic Road Tracking Using Parallel Angular Texture Signature

1021

J. Shen received the Ph.D. degree in cartography and geographic information system Wuhan University, in 2009. She is currently a postdoctoral in Chinese Academy of Surveying and Mapping. Her research interests include airborne LiDAR data processing and information extraction, data mining from spatial data warehouse.

Y. Liang, is a professor of Shandong Agricultural University. His main research interests include digital agriculture, digital water conservancy and digital city. He coordinated more than 10 national and provincial research projects, and has published more than 50 national, international journals papers and 5 books.