On-Road Vehicle Recognition Using the Symmetry Property and Snake Models

Abstract

Vehicle recognition is a fundamental task for advanced driver assistance systems and contributes to the avoidance of collisions with other vehicles. In recent years, numerous approaches using monocular image analysis have been reported for vehicle detection. These approaches are primarily applied in motorway scenarios and may not be suitable for complex urban traffic with a diversity of obstacles and a clustered background. In this paper, stereovision is firstly used to segment potential vehicles from the traffic background. Given that the contour curve is the most straightforward cue for object recognition, we present here a novel method for complete contour curve extraction using symmetry properties and a snake model. Finally, two shape factors, including the aspect ratio and the area ratio calculated from the contour curve, are used to judge whether the object detected is a vehicle or not. The approach presented here was tested with substantial urban traffic images and the experimental results demonstrated that the correction rate for vehicle recognition reaches 93%.

Keywords

Contour Extraction Vehicle Recognition Symmetry Snake Model Stereovision

1. Introduction

Vehicle recognition is a fundamental task for on-board driver assistance systems, enabling cars to be aware of their driving environment and warning drivers of potential hazards. Vehicle recognition has many applications, including platooning, stop and go, autonomous navigation and parking optimization, etc. In the past decade, extensive research has been carried out in the field of vehicle detection using on-board imaging systems together with image processing techniques.

Vision-based vehicle recognition methods normally follow two basic steps [1]: 1) hypothesis generation (HG), which hypothesizes the regions in an image where possible vehicles are; and 2) hypothesis verification (HV), which verifies the correctness of the hypothesis in step 1) by using certain algorithms. In general, HG is based on some simple features. There are three kinds of methods for locating potential vehicles [1, 2]: 1) knowledge-based methods; 2) motion-based methods; and 3) stereovision-based methods. Knowledge-based methods make use of prior knowledge to locate candidate vehicles in an image. Such prior knowledge includes symmetry [3], colour [4], shadow [5], geometrical features [6] and texture [7], etc. As one of the main signatures of man-made objects, symmetry has been used for object detection and recognition in computer vision [8]. In [9, 10], symmetry properties were used to filter out homogeneous areas and segment potential vehicles where the images of vehicles observed from rear or frontal views are in general symmetrical in the horizontal and vertical directions. In [3], the detection process uses shadow and symmetry features to generate vehicle hypotheses. In [11], vehicle symmetrical contours and license plate positions were used to detect vehicles. However, these studies were only suitable for detecting vehicles in the same lane whose rear view image strictly conforms to the symmetry property. Moreover, symmetry estimations are sensitive to noise for homogeneous areas. In general, knowledge-based methods are effective in relatively simple environments but cannot work in complex environments because the prior knowledge is susceptible to illumination, image quality and the complexity of background. Motion-based methods commonly use optical flow to estimate the relative motion of vehicles. The disadvantages of this method are time-consuming and cannot detect static obstacles. Stereovision-based methods have been employed extensively for HG. Kunsoo Huh [12] proposed a vehicle detection method using stereovision. This method can effectively and accurately locate a potential vehicle, but is unable to determine whether it is a vehicle or not.

Approaches using HV are classified into two categories: 1) template-based methods; and 2) appearance-based methods. Template-based methods use predefined patterns of the vehicle class and perform correlations between the image and the template. The templates are commonly proposed based on the vertical symmetry of the frontal/rear view of the vehicle. Since a vehicle's appearance changes with the viewing angle, it is difficult to build a vehicle class template. Moreover, the templates are sensitive to image noise. Appearance-based methods learn the characteristics of the vehicle class from a set of training images that should cover any variability in a vehicle's appearance. These methods usually contain the following steps: 1) train a classifier, such as a support vector machine (SVM) [13], a neural network [14] or the Adaboost method [15], with substantial data 2) extract features from the object, such as Haar [16] features or HOG [17] features, and then input these features into the classifier to identify these objects. In general, appearance-based methods are more accurate than template-based methods; however, they are more costly due to the need for classifier training.

In this paper, we propose a novel vehicle recognition approach that identifies a vehicle based on the object geometry contour shape. The approach integrates stereovision with a novel object contour extraction algorithm. First, we locate potential vehicles by the stereovision-based method described in our previous work [18, 19]. Second, an object contour curve is extracted using symmetry properties and refined by the snake model technique. Afterwards, two factors including the object aspect ratio and the area ratio calculated from the contour curve are used to identify vehicles.

The contributions of this work can be summarized as follows: 1) using the snake model together with the object symmetry property for the extraction of a complete obstacle's contour curve is a novelty. As a contour extraction technique, the snake model has been widely used in many image analysis applications with relatively simple background information. However, no work has been reported on using the snake model for complex road scene image analysis; 2) integrating stereovision with the snake model overcomes a key limitation of the snake model in that the snake model is sensitive to the initial position of the contour curve. Stereovision can work perfectly well in locating the obstacle region; 3) the approach developed here not only works for direct rear/front view vehicles but also for vehicles observed from the side view, which is a significant advance in comparison with existing studies; 4) a contour curve is the most straightforward cue for obstacle recognition. Therefore, a contour extraction technique suitable for clustered and dynamic traffic image analysis is crucial for obstacle detection and classification in advanced driver assistance systems. The approach proposed here can be extended for multi-class object recognition tasks; 5) the approach proposed here divides obstacle recognition into two steps - namely object detection and classification - which differ from most existing object recognition applications.

2. Approaches

2.1 Overall Structure

The schematic diagram of the proposed system is illustrated in Fig. 1. Firstly, a stereovision rig corroborated with an edge-indexed stereo matching algorithm [18] is employed to produce an edge-based disparity map. The regions of interest (ROIs) in the left image are created according to disparities between the objects and they comprise the potential vehicle objects in the scene. Secondly, symmetry properties are applied to eliminate the non-boundary points so that an initial object contour can be extracted. Subject to noise, the initial contour obtained can be discontinuous and incomplete, containing some noise points within the vehicle body. Therefore, the initial contour is refined with the snake model so that a closed and complete object contour curve is generated. Thirdly, two parameters including an object aspect ratio and an area ratio are calculated from the contour curve and accordingly the objects within the ROIs are classified into two classes - vehicles or other objects.

Figure 1.

The schematic diagram of the vehicle recognition system

2.2 Object Segmentation based on Stereovision

The stereovision technique is an imitation of the physiology of the human eye. The stereovision technique has been successfully used for object segmentation, 3-D reconstruction and target tracking, etc. For a vehicle recognition system, one of the most important and difficult processes is to segment the potential vehicle from the complex traffic scene. In this paper, we segment the objects by the following steps: 1) generate a vertical edge image for the left image using a Sobel operator; 2) construct the disparity map using the edge-indexed stereo matching method; 3) construct the depth image and segment the objects according to their distance; 4) reconstruct the segmented objects into the left edge image and remove the noise points based on a disparity threshold. The details of the approach can be found in [18, 19]. The edge-indexed stereo-matching loses some image information, but it greatly reduces the computational cost in comparison with a dense stereo-matching algorithm. Therefore, it is suitable for our application in locating a region of a potential vehicle subject to a real-time running requirement.

2.3 Contour Extraction using Symmetry Properties

2.3.1 Generalized Symmetry Properties

Symmetry is an inherent properties of most man-made objects, such as vehicles, houses, windows, tables, etc. They are normally symmetrical about the vertical axis. However, the symmetry of an object also depends on the angle of viewing. In this application, only an exact rear view or front view of a vehicle satisfies the symmetry relation.

Fig. 2(a) is the contour curve for a vehicle with a direct rear-view while Fig. 2(b) shows the same vehicle with a viewing angle, where the blue curve is the contour for the vehicle with a direct rear-view and the yellow dashed-curve is the contour with a viewing angle. It can be seen from Fig. 2(a) that the contour curve is symmetrical about the axis X-X′. However, it is no longer symmetrical about the axis X-X′ for the yellow contour curve in Fig. 2(b) (i.e., points a and b′ no longer meet equation: x′_i+x_r=2x_a, where X_a is the abscissa of the symmetry axis X-X′, and x_l and x′_r are the abscissa for points a and b′ respectively). Instead, these points meet the following relation:

Figure 2.

Definition of generalized symmetry

x_{l i} + x_{r i} \pm ε_{i} = x_{a}

(1)

where ɛ_i is a value representing the degree of asymmetry between x_li and x_ri about x_a. The subscripts l, r and i represent left, right and the ith point, respectively. By defining a threshold ɛ_th, the points that meet an inequality ɛ_i ≤ ɛ_th can be regarded as a generalized symmetrical point pair. That is, equation (1) can be rewritten to represent a generalized symmetrical point pair as follows:

| x_{l i} + x_{r i} - x_{a} | \leq ε_{t h}

(2)

The threshold ɛ_th reflects the degree of asymmetry between the left and right points of a point pair. In Fig. 2(a), the vehicle contour is almost strictly symmetrical about the axis X-X′; therefore, ɛ_th can be set as zero. In Fig. 2(b), the yellow contour is not strictly symmetrical any more due to the side-view effect - this is the reason why we need extend symmetry to generalized symmetry. Actually, we hope that ɛ_th is as small as possible; thus, we try to move the axis X-X′ to the axis Y-Y′ to achieve a relatively-strict symmetry.

2.3.2 Initial Contour Image Extraction based on Symmetry Properties

The stereovision approach locates a region containing potential vehicles. The next step is to extract the contour of a vehicle by using its symmetry properties. When extracting an initial contour image, we are only concerned with the vertical edges of the objects segmented by the stereovision approach. The principle of extracting the initial contour of an object is based on the fact that the edge point pairs on the contour satisfy the inequalities (2). Consequently, other edge points inside or outside the vehicle body are regarded as noise points.

The initial contour image extraction using object symmetry properties is as follows:

Step 1 . Estimating the initial symmetry axis: within the vertical edge image, the initial symmetry axis A_i, is determined as a vertical line whose abscissa is the mean value of the abscissa X_m of all the edge points.

Step 2 . Estimating the real symmetry axes: calculate the mean square deviations ${\bar{σ}}_{l e f t}$ , ${\bar{σ}}_{r i g h t}$ , of the abscissa of both side edge points divided by A_i, which reflect the dispersion of the edge points. If ${\bar{σ}}_{l e f t} > {\bar{σ}}_{r i g h t}$ , this means that the real symmetry axis should be on the left-hand side of A_i, otherwise, it would be on the right-hand side.

Step 3 . Extracting the initial contour using symmetry properties: move A_i towards the left or the right, horizontally within a range of Δa, to search for the true symmetry axis. The moving range Δa is determined according to the difference of ${\bar{σ}}_{l e f t}$ and ${\bar{σ}}_{r i g h t}$ . Each time, move A_i with a small step size Δs (e.g., three pixels) and search for symmetry points according to equation (2) about the symmetry axis and the search number n = Δa/Δs. The true symmetry axis is the one that has the maximum symmetry point pairs. Remove any asymmetrical point pairs within the vertical edge image and get the initial contour of the vehicle.

Step 4 . Repairing the initial contour: connect the top and bottom horizontal symmetry point pairs.

2.4 Contour Refinement using the Snake Model

Snake models (also called ‘active contour models’) were proposed by Kass et al. [20] as a segmentation scheme and have been successfully applied in a variety of problems in computer vision and image analysis, such as edge and subjective contour detection [21], motion tracking and segmentation [22]. It has become a main method for contour extraction.

A snake curve can be construed as a number of control points that are linked together to form a contour and deform under the constraining forces. The deformation is carried out by minimizing an energy function so that the contour will move from an initial position to the true contours of objects. A classic snake model is defined as a curve C(s) = [x(s, t), y(s, t)], s ∈ [0, 1], where [x, y] is a point in the image that moves through the spatial domain of the image to minimize the energy function E(C):

E (C) = \int_{0}^{1} \frac{1}{2} [α {| C^{'} (s, t) |}^{2} + β {| C^{″} (s, t) |}^{2}] + E_{e x t} d s

(3)

where α and β are weighting parameters that control the snake's tension and rigidity, respectively, and C'(s) and C“(s) denote the first and second derivatives of C(s) with respect to s. The external energy E_ext is derived from the image so that it takes on its smaller values at the features of interest, such as boundaries. A snake that minimizes E(C) must satisfy the dynamic equation:

C_{t} (s, t) = α C ″ (s) - β C ″ ″ (s) - \nabla E_{e x t} = 0

(4)

The curve keeps moving towards the contour of the object and forms a closed parametric boundary curve until the motion stops. However, there are several shortcomings of this model: 1) it is in general sensitive to the initial position of the contour curve (i.e., the initial contour must be close to the true boundary or else it will likely converge on a wrong result); 2) the running speed of it is very slow because the Scope of the external force field F_ext is limited; and 3) it does not have the ability to repair the boundaries with gaps.

In this work, the first shortcoming is overcome by using stereovision, since the stereovision approach generates a ROI, which minimizes the object region. The initial position of the contour curve is set in the ROI, which is very close to the actual position of the object under detection. In order to overcome the other two shortcomings, we use an improved snake model called the ‘distance potential model’ proposed by Cohen [23]. The distance potential model defines E_ext as a function of distance. The potential force function can be expressed as:

F {(x, y)}_{d i s \tan c e} = f (d (x, y))

(5)

where d(x, y) is the Chamfer distance or Euclidean distance between the image point (x, y) and the nearest boundary. In our work, f(d(x, y)) is defined as the negative gradient of d(x, y). This improved snake model effectively expands the range of external force and improves the ability to extract a contour for an edge with gaps.

2.5 Vehicle Recognition based on Object Contour

After the object contour is extracted, the aspect ratio R_th and the area ratio R_a are calculated from the contour and used for vehicle recognition.

2.5.1 The Aspect Ratio R_hw of the Contour Curve

A vehicle is a rigid object and has a fixed aspect ratio (i.e., height/width ratio), which is significantly different from other road obstacles such as pedestrians. Normally, a vehicle has a significantly smaller aspect ratio than a pedestrian. This property can be used for recognizing a vehicle. However, the aspect ratio of a vehicle may vary with the angle of view. Fig. 3 shows the relative position between a vehicle-mounted camera and observed objects in a traffic scene. For two vehicles with the same size in Pos1 and Pos2 (the same distance), the width observed by the camera is different. The vehicle width W′ in Pos2 is L.tanβ + W, where W is the width of the vehicle in Pos1. The aspect ratio R_hw of a vehicle in different positions satisfies the following inequality:

Figure 3.

Vehicle-mounted camera observation geometry and the contour curve of vehicles: (a) the aerial view of the pavement; (b) the contour curve of the vehicle located in lane two; (c) the contour curve of the vehicle located in lane three.

\frac{H}{W} \leq R_{h w} \leq \frac{H}{L . \tan β + W}

(6)

where 2β is the horizontal viewing angle of the camera and H is the vehicle height. Actually, R_hw also varies with the vehicle class.

2.5.2 The Area Ratio R_a of the Contour Curve

Normally, a vehicle has a significantly smaller aspect ratio than a pedestrian. However, there are some exceptional cases. Fig. 4 shows one of these cases, where a pedestrian is swinging his/her arms.

Figure 4.

The ratio of area for the contour

In this case, the contour curve of the pedestrian has a smaller R_hw and may be wrongly regarded as a vehicle according to R_hw. It can be seen from Fig. 4 that the actual area of the pedestrian contour curve is much less than the nominal area H×W - that is, the ratio of the actual area of the contour to H×W of a pedestrian is significantly smaller than that of a vehicle. The ratio of the actual area of the contour to H×W is called the ‘area ratio’ and can be expressed as:

R_{a} = \frac{\int f (x, y) d s}{H \cdot W}

(7)

where the integral (i.e., the numerator) is the actual area of the contour. The significant difference of the area ratio between a vehicle and a pedestrian is to be used as a further clue in recognizing a vehicle.

3. Experiments and Results

3.1 Stereovision-based Object Segmentation

Fig. 5(a) is the left image of a stereo image pair for a typical traffic scene to be analysed in this paper. Fig. 5(b) is the disparity image generated from the edge-indexed stereo matching, in which the colours of the point denote the disparity value. The edge image is generated by a Sobel operator only for the vertical direction. Fig. 5(c) is the depth map (a bird's eye view) using disparity information in the fig. 5(b) and the object segmentation results in the depth map. Fig. 5(d) is the corresponding object segmentation resulting in the disparity map.

Figure 5.

The process of object segmentation: (a) the left image of the stereo image pair; (b) the disparity image; (c) the depth map with a lateral range from −8 m to 8 m, a distance range from 4 m to 40 m, a range resolution of 0.2 m × 0.2 m, and points on the road surface, have been discarded; (d) the segmentation objects are outlined with a red rectangle.

3.2 Contour Extraction and Refinement using the Symmetry Properties and the Snake Model

Fig. 6 shows the process of contour curve extraction for Obj3 by the symmetry properties and the snake model.

Figure 6.

The process of contour curve extraction for Obj 3: (a) the original image of Obj3; (b) the vertical edge image; (c) the boundary contour image and the symmetry axis; (d) the initial contour; (e) the iterative trajectory of the snake model; (f) the refined contour curve of Obj 3.

Fig. 6(a) is the original image of Obj3 located in the adjacent lane. Fig. 6(b) shows the vertical edge image of Obj3 segmented using the stereovision-based method. The symmetry axis and boundary contour image extracted by the general symmetry properties are shown in Fig. 6(c), where the green dashed line is the initial symmetry axis and the red line is the true symmetry line. Fig. 6(c) is the result generated until step 3 in section II-C-2. Fig. 6(d) shows the initial contour image of Obj3 after connecting the top and bottom horizontal symmetry point pairs (step 4). It can be seen that the initial contour is discontinuous and contains some noise points that do not lie in the object boundary. Therefore, the snake model is applied to refine the initial contour so that a closed contour can be generated, as shown in Fig. 6(f). Fig. 6(e) shows the process of extracting the contour curve by the snake model, where a series of red curves are the iterative trajectory for the initial curve and the curve moves towards the boundary continuously under the action of external forces.

Fig. 7 shows the process of contour curve extraction for Obj2. The results for each stage are the same as for Fig. 6. It can be seen that the initial symmetry axis (green) and the true symmetry axis (red) are almost coincident for Obj2 and are far apart from each other for Obj3. This is because Obj2 is a direct rear image with a strict symmetry while Obj3 is observed with a side view angle resulting in a relatively loose symmetry.

Figure 7.

The process of contour curve extraction for Obj2: (a) the original image of Obj2; (b) the vertical edge image; (c) the boundary contour image and symmetry axis; (d) the initial contour of Obj2; (e) the iterative trajectory of the snake model; (f) the refined contour curve of Obj2.

In order to verify the generality of the approach, multiple classes of vehicles are analysed in this work. The processes of contour curve extraction for different classes of vehicles are shown in Fig. 8 and Fig. 9. Fig. 8 shows the process of contour curve extraction for a small car. Here, Fig. 8(a) presents a slight side view for a sedan located in the adjacent lane. Fig. 8(b) ∼ Fig. 8(d) are the direct rear views for moving vehicles located in the same lane. Fig. 8 (a1) ∼ Fig. 8 (d1) show the corresponding symmetry axis and the boundary contours. Fig. 8(a2) ∼ Fig. 8(d2) are the extracted contour curves.

Figure 8.

The process of contour curve extraction for small cars: (a) the original image for a moving car located in an adjacent lane; (b) ∼ (d) the original image for small cars with a direct rear-view; (a1) ∼ (d1) the corresponding boundary contours and symmetry axis; (a2) ∼ (d2) the extracted contour curves.

Figure 9.

The process of contour curve extraction for mid-sized vehicles and larger vehicles: (a), (b) the original image for a moving bus located in an adjacent lane; (c) the original image for a mid-sized bus with a direct rear-view; (d) the original image for a truck with a direct rear-view; (a1) ∼ (d1) the corresponding boundary contours and symmetry axis; (a2) ∼ (d2) the extracted contour curves.

Fig. 9 shows the process of contour curve extraction for mid-sized vehicles and larger vehicles. Fig. 9(a) and Fig. 9(b) present slight side views of buses. Fig. 9(c) and Fig. 9(d) show the bus and truck located in the same lane as the observing vehicle.

From the results shown in Fig. 6(c), Fig. 8(a2), Fig. 9(a2) and Fig. 9(b2), it can be seen that the initial contour can be extracted successfully for a vehicle located in the adjacent lane even though their appearance is no longer strictly symmetrical. For all of them, the initial symmetry axis is far from the true symmetry axis because of the asymmetry of the side view shapes of the vehicles. However, the initial symmetry axis almost overlaps with the true symmetry axis for all the direct rear view images. This proves that the method proposed here is effective for extracting the true symmetry axis and object contour curves.

3.3 Vehicle Recognition

Fig. 10 shows some of the standard vehicle samples and their contour curves. It can be seen that the R_hw of a sedan is smaller than that of a mid-sized vehicle and a larger vehicle. The larger the vehicle is, the larger the R_hw is. For the same class of vehicle, the one located in the adjacent lane has a smaller R_hw than that located in the same lane. Therefore, we can calculate R_hw1 and R_hw2 according to the size of the vehicle, the viewing angle β and equation (6), in order to determine a range of the R_hw of the vehicle class. A comprehensive analysis of different types of vehicles and viewing angle factors give the result that R_hw must be within 0.4 ∼ 1.4 for a vehicle-type object. Furthermore, R_a is set to be greater than 0.7 for a vehicle-type object. In this work, a detected object is classified into two types - vehicle or non-vehicle.

Figure 10.

Vehicle samples and their contour curve: (a) a small car; (b) a mid-sized vehicle; (c) a larger vehicle; (d) an oblique view of a small car located in an adjacent lane; (e) ∼ (h) the actual contour curves of the vehicle samples.

In this work, a total of 200 obstacles captured in different driving scenarios from a direct front/rear view are tested, and all of their contour curves can be extracted successfully. The verification results are as follows.

Table 1 shows the verification results only using R_hw as a criterion.

Table 1.

Recognition results using R_hw

Objects	Vehicles			Non-vehicle
Objects	Small car	Mid-sized vehicle	Large vehicle	Non-vehicle
Actual number	60	52	44	44
Correct number	56	47	40	31
Correction rate (%)	93.3%	90.4%	88.6%	70%
Correction rate (%)	91.7%			70%

Table 2 shows the verification results only using R_a as a criterion.

Table 2.

Recognition results using R_a

Objects	Vehicles			Non-vehicle
Objects	Small car	Mid-sized vehicle	Large vehicle	Non-vehicle
Actual number	60	52	44	44
Correct number	52	43	38	39
Correction rate (%)	86.7%	82.7%	86.3%	88.7%
Correction rate (%)	85.2%			88.7%

It can be inferred from Table 1 and Table 2 that the correction rate for positive samples (vehicles) using R_hw is higher than that using R_a, but it is lower for negative samples (non-vehicles).

Table 3 shows the verification results using both R_hw and R_a.

Table 3.

Recognition results using R_a and R_th

Objects	Vehicles			Non-vehicle
Objects	Small car	Mid-sized vehicle	Large vehicle	Non-vehicle
Actual number	60	52	44	44
Correct number	57	49	41	41
Correction rate (%)	95%	94.2%	93.1%	93.1%
Correction rate (%)	94.2%			93.1%

As shown in Table 3, the correction rate is much improved for both positive samples and negative samples using both R_a and R_hw. Table 4 show the results for vehicles located in the adjacent lane.

Table 4.

Recognition results for vehicles in the adjacent lane using Ra and R_hw

Objects	Vehicles			Non-vehicle
Objects	Small car	Mid-sized vehicle	Large vehicle	Non-vehicle
Actual number	30	22	18	13
Correct number	27	19	14	14
Correction rate (%)	90%	86.4%	77.8%	92.9%
Correction rate (%)	85.7%			92.9%

As shown in Table 4, the correction rate for vehicles in an adjacent lane is slightly smaller than that for vehicles in the same lane. However, it is still as high as 85.7%. The correction rate for ‘Non-vehicle’ objects in adjacent lanes is 92.9% and does not represent a significant difference to those in the same lane. This is because ‘Non-vehicle’ objects like pedestrians or traffic posts do not have a significant ‘thickness’; therefore, their contours do not make a significant difference, whether they are in the same lane or not.

The results show that R_hw and R_a are useful criteria for vehicle recognition. R_hw has an advantage in the identification of positive samples and R_a is more suitable for identifying non-vehicles. However, the correction rate in integrating both R_hw and R_a is much improved for both positive samples and negative samples. Moreover, the correction rate for small cars is higher than that for mid-and large-sized vehicles.

4. Conclusions

This paper presents a vehicle recognition approach for intelligent vehicles. The approach identifies a vehicle based on the object geometry contour shape. The approach integrates stereovision with a novel object contour extraction algorithm. Stereovision is used to segment objects from clustered road scene images and locate an object region. An object contour is initially extracted by using a vehicle symmetry property and is further refined by a snake model algorithm. An improved snake model - a potential distance model - is adapted and the potential force function is defined as the negative gradient of the Euclidean distance between the image point and the nearest boundary. This snake model effectively expands the range of external force and improves the ability to extract contours for edges with gaps. After the object contour is obtained, two shape factors including an aspect ratio and an area ratio are calculated and used for the classification of the object detected. The approach is verified by substantial scenario images. The experimental results demonstrate that the system is able to successfully track and classify most vehicles observed from the rear/front view and the side view. For a total of 200 direct rear/front view obstacles captured in different driving scenarios, 94.2% of the vehicles were correctly identified and 93.1% of non-vehicles were correctly identified. In addition, the correction rate for vehicles in adjacent lanes was 85.7%. The system was implemented on a normal Pentium dual-core 3.0 GHz PC at a processing rate of 8 frames/sec. The main computation cost comes from the snake model algorithm. However, the system speed could be greatly improved by using special image processing chips, which will constitute our future work.

Footnotes

5. Acknowledgements

This work was sponsored by the National Natural Science Foundation of China (Project No. 61374197), the Science and Technology Commission of Shanghai Municipality (Project No. 13510502600) and supported by the Programme for Professors of Special Appointment (Eastern Scholar) at the Shanghai Institutions of Higher Teaming.

References

Sun

Bebis

Miller

(2006) On-road Vehicle Detection: a Review, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 5, pp. 694–711.

Zhou

Duan

(2012) Machine-vision Based Preceding Vehicle Detection Algorithm: a Review. Proceedings of the 10th world congress on intelligent control and automation, pp. 4617–4622.

Hoffmann

Dang

Stiller

(2004) Vehicle detection fusing 2D visual features. Proceedings of IEEE Intelligent Vehicles Symp., Parma, Italy, pp. 280–285.

Xiong

Debrunner

(2004) Stochastic car tracking with line-and color-based features. IEEE Trans. Intell. Transport. Syst., vol. 5, no. 4, pp. 324–328.

Zhang

(2010) Research on Front Vehicle Detection and Tracking based on Multiple Features. Jiangsu University pp. 26–30 (in Chinese).

Mei

Zhoug

Kevin (2007) Integrated detection, tracking and recognition for IR video-based vehicle classification. Journal of Computers, vol. 2, no. 6, pp. 1–9.

Xia

(2011) Adaptive detection of moving vehicle based on on-line clustering. Journal of Computers, vol. 6, no. 10, pp. 2045–2052.

Marola

(1989) Using Symmetry for Detecting and Tocating Objects in a Picture. Computer Vision, Graphics and Image Processing, vol. 46, pp. 179–195.

Kuehnle

(1991) Symmetry-based Recognition for Vehicle Rears. Pattern Recognition Letters, vol. 12, pp. 249–258.

10.

Bertozzi

Broggi

Fascioli

, (2000) Vision-based Intelligent Vehicles: State of the Art and Perspectives. Robotics and Autonomous Systems, vol. 32, pp. 1–16.

11.

Lian

Zhao

Zhang

(2012) Vehicle Detection based on Information Fusion of Vehicle Symmetrical Contour and License Plate Position. Journal of Southeast University, vol. 28, no. 2, pp. 240–244.

12.

Huh

Park

(2008) A Stereo Vision-based Obstacle Detection System in Vehicles. Optics and Tasers in Engineering, vol. 2008, pp. 168–178.

13.

(2007) Research on Optical Flow Computation for Miton Image Analysis. Xi'an: Xidian University (in Chinese).

14.

Lan

Zhang

(2010) A New Vehicle Detection Algorithm for Real-time Image Processing System. IEEE International Conference on Computer Application and System Modeling, pp. 101–104.

15.

Xia

(2011) Moving Object Classification Method based on SOM and K-means. Journal of Computers, vol. 6, no. 8, pp. 1654–1661.

16.

Sayanan

Trivedi

M. M.

(2009) Active Tearning-based Robust Monocular Vehicle Detection for On-road Safety Systems. IEEE Intelligent Vehicles Symposium, pp. 399–404.

17.

Mao

Xie

Huang

(2010) Preceding Vehicle Detection Using Histograms of Oriented Gradients. International Conference on Communication, Circuits and systems, pp. 354–358.

18.

Huang

(2005) Obstacle Detection in Urban Traffic using Stereo-vision. IEEE International Conference on Intelligent Transportation Systems, pp. 13–16, Vienna, Austria.

19.

Huang

Thompson

(2005) Stereovision-based Object Segmentation for Automotive Applications. EURASIP Journal on Applied Signal Processing, vol. 14, pp. 2322–2329.

20.

Kass

Witkin

Terzopoulos

(1987) Snakes: Active Contour Models. International Journal of computer vision, vol. 1, no. 4, pp. 321–331.

21.

Chen

Lai

(2007) Active Contour Models on Image Segmentation: a Survey. Journal of Image and Graphic, vol. 1, no. 12, pp. 11–20.

22.

Michal

Ohad

B. S.

(2011) Free Boundary Conditions Active Contours with Applications for Vision. International Conference on ISVC, pp. 180–191.

23.

Cohen

T. D.

Cohen

(1993) Finite-element Methods for Active Contour Models and Balloons for 2-D and 3-D images. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15, no. 11, pp. 1131–1147.

On-Road Vehicle Recognition Using the Symmetry Property and Snake Models

Abstract

Keywords

1. Introduction

2. Approaches

2.1 Overall Structure

2.2 Object Segmentation based on Stereovision

2.3 Contour Extraction using Symmetry Properties

2.3.1 Generalized Symmetry Properties

2.3.2 Initial Contour Image Extraction based on Symmetry Properties

2.4 Contour Refinement using the Snake Model

2.5 Vehicle Recognition based on Object Contour

2.5.1 The Aspect Ratio Rhw of the Contour Curve

2.5.2 The Area Ratio Ra of the Contour Curve

3. Experiments and Results

3.1 Stereovision-based Object Segmentation

3.2 Contour Extraction and Refinement using the Symmetry Properties and the Snake Model

3.3 Vehicle Recognition

4. Conclusions

Footnotes

5. Acknowledgements

References

2.5.1 The Aspect Ratio R_hw of the Contour Curve

2.5.2 The Area Ratio R_a of the Contour Curve