2.1 Calibration Methods
The Camera calibration is a major work in computer vision, robotics, computer-integrated manufacturing systems, and automation. Camera calibration is the estimation and determination of some parameters necessary to establish projection equations between world coordinates and pixel coordinates. Camera self-calibration techniques make it possible to compute three-dimensional world coordinates of observed scene from video sequences without getting knowledge of the camera intrinsic parameters previously. The notion of self-calibration was introduced by Maybank and Faugeras for the perspective camera and by Hartley for the rotating camera. A self-calibration method for affine camera was given by Long. Over the past few decades, considerable endeavor has been made on development of effective and accurate procedures and algorithms to transfer the pixel coordinates to world coordinates, that is to say, identify the internal camera geometric and optical characteristics (intrinsic parameters) or the three-dimensional position and orientation of the camera frame relative to a certain world coordinate system (extrinsic parameters). In our study, camera calibration is based on the vanishing point of two lines known to be parallel on the road surface. Most work in camera calibration involves calibrating both intrinsic and extrinsic camera parameters simultaneously. However, real-time information is critical in many situations for successful applications. Accordingly, an efficient extrinsic parameters calibration algorithm that requires less preparation in manually operation setup and less computation in processing would be extremely useful.
Most calibration methods [12]-[15] use known features in a scene to evaluate the camera parameters, in [12], the calibration procedure is based on using the Camera Calibration
Toolbox for Matlab [24]. A calibration equation that separates rotational and translational parameters is given in [13]. In [13], a four-point calibration procedure is proposed that includes three points on a line and one point out of the line, and results in four possible calibration solutions, obtained analytically. For the reason of simplicity, in [13], it has been assumed that the camera world coordinates and the image coordinate axes are identical, and Z and z are the optical axis of the camera. In [14], it has presented a camera self-calibration technique which is able to handle time-varying focal length change. It has been exhibited that this technique can track focal length change even under relatively noisy conditions; however, it still uses known features in a scene to evaluate the camera parameters. In [16] and [17], sets of parallel lines of a hexagon are employed to calculate the camera parameters. Results from these papers show that parallel lines can be used to appropriately determine the camera parameters. Efficient algorithms [18]-[20] have been developed for evaluating the camera parameters utilizing parallel lanes in a traffic scene. In [18], a calibration algorithm for a single camera overlooking traffic is proposed. That algorithm needs the technician installing that system, measuring the height and tilt of the camera, and selecting the road edges in an image. That is to say, their method needs particular manual operations to measure the tilt angle of the camera. It shows how the camera calibration parameters can be computed from the height, the tilt, and the four selected points. Multiple parallel lanes and a particular perpendicular line were used to calibrate the camera parameters in [19] and [20]. Another method is to use special calibration targets to substitute for measurement of corresponding points. In this case; however, the calibration target must still be carefully installed relative to the world frame.
In this paper, a novel equation will be derived to calculate the vanishing point from two lines we give in pixel coordinate. The derivation requires only a single set of parallel line, the reference width, and the reference height. Compare the proposed method with the existing
approaches and we’ll see our method has the advantage of requiring neither the camera’s information nor multiple sets of parallel lines.
2.2 AdaBoost Algorithm’s Application
In P. Viola and M. J. Jones [3] proposed an original feature selection scheme for human face detection. The algorithm uses training data to comprise a cascade of boosted classifiers, each layer in the cascade classifier rejects some input and regions that do not contain interested object. This algorithm uses Haar-like features, also called rectangular filters (confirmed by C. Papageorgiou et al. [25]), and the training algorithm is AdaBoost machine learning algorithm [1], which is applied to selecting a number of useful features in each layer. The application of integral image is to accelerate the calculation of Haar-like features.
The characteristic of cascaded structure makes the classifier trained by AdaBoost suitable for real-time face-detection application, and this approach has inspired a lot of recent works in vehicle detection.
Inspired by AdaBoost algorithm, P. Negri et al. [26] combined the rectangular filters (Haar-like features) and the histograms of oriented gradient (HoG) with AdaBoost algorithm.
The fusion takes advantage of the two detectors: generative classifiers composed of Haar-like features filter out easy negative inputs in the early layers of the fusion classifier. While in the later part of the fusion classifier, the discriminative classifiers composed of HoG features generate a fine decision boundary to remove the negative inputs which are similar to vehicle, so that the fusion achieved better performances than either feature.
In this study, we use cascaded classifier trained by AdaBoost algorithm to generate a vehicle detector. We combine the tracking algorithm of vehicle’s trajectory we proposed with AdaBoost vehicle detector to provide information for further calibration.