1.1 Motivation
As the conveyances are getting growth with years, the traffic is becoming more and more serious in most developed countries. A lot of researches about the intelligent transportation systems (ITS), including the smart vehicles, the driving safety, and the traffic mobility, have been proposed in recent years. In fact, many problems are still expected to be overcome.
Above all, one of the most interesting and important issues for the ITS applications is concerning the smart vehicles.
It is necessary to acquire the information about the on-road obstacles and the lane tendency while driving on the way. Thanks to the driver’s careless attitude, his/her moving vehicle may hit the obstacles on the road, or may deviate from the correct lane orientation, which induces the traffic accidents. Hence the on-vehicle obstacle and lane detection system plays a fundamental and essential role in moving vehicles. Such a system can either be the driver assistance function to warn the drivers of occurrences of which they may not be aware, or be the vision system of unmanned vehicles to supply the car controller with the road information for the goal of the automatic driving.
In general, the vision-based obstacle and lane detection system is a good choice for ITS applications. Cameras are mounted on the vehicle, and then the road images are captured and processed. The systems based on the vision have advantages of the high spatial resolution and the fast image scansion. Many approaches using the image processing have been developed [1], and different techniques will be reviewed in the next section.
1.2 Background
1.2.1 Related work of obstacle detection
The definition of obstacles induces the development of detection algorithms. Since the vehicles are most of obstacles on the road, some approaches to detect obstacles are limited to search for particular features and then to match them with specific patterns, such as the symmetry, textures, shapes, an approximate contour, and so on. In this case the processing can be focused on the analysis of a single still image. Broggi et al. perform a function of vehicle detection to locate and track the vehicle by exploiting the symmetry of the rear parts of a typical car and a bounding box satisfying specific aspect ratio constraints [2]. However, such a pattern-based approach may fail when characteristics of obstacles do not match the pre-defined model.
As we know, vehicles are not the only obstacles on the road. A generic obstacle is defined as an object rising out significantly from the road surface. Following this definition, the pattern-based approach does not work owing to the lack of a prior knowledge about generic obstacles on the road. More complex techniques must be imported to handle such a problem, and two and more images may need to be taken into account.
The optical flow-based approach utilizes a sequence of two or more images to obtain reliable and dense optical flows. In the assumption of the small difference between two successive images due to the short time interval, the two-dimensional motion between two images approximates the single direction. And therefore, the optical flow field can be computed and the ego-motion can be estimated. Giachetti et al. use a correlation technique to compute the flow field, and the obstacles moving with different speeds can be segmented by analyzing the velocity fields [3]. However, the optical flow-based approach may fail deriving from the lack of textures on the road, or from large displacements between two consecutive
frames due to the higher speed or vibrations of the vehicle.
Another technique similar to the optical flow-based approach is known as the motion-based method by estimating the motion of the ground plane and then detecting the obstacles whose motions differ from that of the ground [4-7]. In this method, it is necessary to make a tracking about the motion among images for large displacements, and as a consequence the assumption of rectilinear motions in optical flow-based methods is invalid.
Since the scenes vary very much among images, it is difficult to identify the pixel correspondence. If the size of searching area is too small, the correct matching for the corresponding pixels may be missed. On the other hand, if the size is too large, too many possibilities may exist. Notice that both optical flow-based and motion-based approaches need expensive computational costs.
The stereo vision-based technique is also used to detect the generic obstacles. The GOLD system transforms both left and right stereo images into top views in order to remove the perspective effect. The ideal square obstacle is transformed into two triangles in the difference image of both remapped views. The polar histogram is constructed from the difference image and then the two peaks in the polar histogram are joined to identify the obstacle [2, 8].
Labayrade et al. also use both left and right stereo images to construct v-disparity image to detect potential obstacles whose disparities differ from that on the road surface. The angles between the cameras and the road are then estimated [9-10]. In conclusion the stereo vision-based method is a better framework than others, and is adopted in this thesis.
1.2.2 Related work of lane detection
It is the objective for lane detection to detect the relative position between the vehicle and the road, and to determine the lane information, such as the offset, the orientation, the curvature, and so forth. Since the structured roads are met in the practical applications, most researches focus on the analysis of marking roads where lane markings are painted on the road surface. Several features of the lane markings, including the constant lane width, the higher brightness on markings, the structured lane shape, etc.
The GOLD system removes the perspective effect by mapping the road image into the top view, and determines the lane markings by relying on the feature of the constant lane width, which may fail when the assumption of a flat road is not valid [8]. Based on the GOLD system, Jiang et al. model the lane as two straight lines to estimate the inclined angle on the condition of non-flat roads [11].
However, the road shape usually is not straight in real cases. Polynomials or splines may be better lane fittings than the straight line. Such a geometric model-based lane detection technique is more robust against the interferences such as shadows, textures, or other vehicles.
Based on the lane geometry, the coefficients of lane model can be found out by several methods. LOIS, LANA, and RVP-I systems decide the coefficients with the maximum likelihood by completely searching the parameter spaces where all possibilities produced in the training phase are built [12-14].
Instead of searching throughout the databases, some road features are detected in the overall image in order to determine the coefficients. Yue Wang et al. [15] and Goldbeck et al.
[16] use the edge-oriented methods to measure the matching degree between the model and the edge map in order to determine the parameters, respectively. Gonzalez et al. classify the objects in the image as the road surface, markings, or obstacles by a histogram-based segmentation method and then pixels belonging to the markings are taken into the fitting of
the lane model [17].
Different from the lane geometry, the statistical model can be used to specify the detection region of interest (ROI) in order to narrow the searching area [7, 18]. On the other hand, the ROI can also be determined according to the features of markings used in the TFALDA [19]. However, the statistical parameters and the weights of marking features have to be trained in advance.
Since the model-based approaches have more robust results and the use of the detection ROI can reduce the computational cost, both ideas are adopted in this thesis. The details will be proposed later.
1.3 Organization
This thesis is organized as follows. A review of algorithms about the obstacle and lane detections is given in this chapter. The preliminary knowledge of the computer vision is introduced in Chapter 2. In Chapter 3, the algorithm of the generic obstacle detection based on two top and bottom stereo cameras is developed. The approach to detect the lane is proposed in Chapter 4. And afterward the experimental results of both obstacle and lane detections are demonstrated in Chapter 5. Finally, a conclusion is presented in Chapter 6.