شناسایی خط خروج بر اساس LBPE، تغییر شکل هاف و رگرسیون خطی
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|24196||2005||25 صفحه PDF||سفارش دهید||7637 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Computer Vision and Image Understanding, Volume 99, Issue 3, September 2005, Pages 359–383
This paper presents a lane-departure identification (LDI) system of a traveling vehicle on a structured road with lane marks. As is the case with modified version of the previous EDF-based LDI approach [J.W. Lee, A machine vision system for lane-departure detection, CVIU 86 (2002) 52–78], the new system increases the number of lane-related parameters and introduces departure ratios to determine the instant of lane departure and a linear regression (LR) to minimize wrong decisions due to noise effects. To enhance the robustness of LDI, we conceive of a lane boundary pixel extractor (LBPE) capable of extracting pixels expected to be on lane boundaries. Then, the Hough transform utilizes the pixels from the LBPE to provide the lane-related parameters such as an orientation and a location parameter. The fundamental idea of the proposed LDI is based on an observation that the ratios of orientations and location parameters of left- and right-lane boundaries are equal to one as far as the optical axis of a camera mounted on a vehicle is coincident with the center of lane. The ratios enable the lane-related parameters and the symmetrical property of both lane boundaries to be connected. In addition, the LR of the lane-related parameters of a series of successive images plays the role of determining the trend of a vehicle’s traveling direction and the error of the LR is used to avoid a wrong LDI. We show the efficiency of the proposed LDI system with some real images.
Machine vision is expected to contribute greatly to the development of vehicle safety and in turn the reduction of traffic accidents caused by driver’s inattention, mistake or lack of experience. However, machine vision has showed limited success because it does not achieve an appropriate reliability in natural outdoor environments. To make machine vision systems practical, robustness is required. This paper presents a machine vision system capable of identifying lane departure (LD) of a traveling vehicle with minimum false alarms and miss-detections on roads with lane marks. The LD identification (LDI) determining whether or not a traveling vehicle deviates from its lane is conducted with lane-related parameters composed of the orientations and positions of lane boundaries in an image. An important observation is that if the optical axis of a CCD camera and the center of lane nearly coincide as shown in Fig. 1A, the angles θl and θr, and distances ρl and ρr, which are defined in Fig. 1B for both of left- and right-lane boundaries, have the values as θl/θr ≈ 1 and ρl/ρr ≈ 1. If the optical axis deviates far from the center of lane, θl/θr and ρl/ρr deviates in accordance from the value of 1. When the optical axis of a CCD camera approaches the left boundary, θl and ρl decrease at the same time while θr and ρr increase. Conversely, when the optical axis approaches the right boundary, θr and ρr decrease and θl and ρl increase. The basic idea of the proposed LDI system is based on this observation. Fig. 2 shows a typical phenomenon where as a vehicle deviates from its traveling lane as shown in Figs. 2A and D, the orientation and position of lane boundaries change at the same time. The two columns on the left of Fig. 2 illustrate the LD to the left side and show that θl and ρl decrease as the vehicle approaches the left-lane boundary as shown in Figs. 2B and E. The two columns on the right depict LD to the right side and show that θr and ρr decrease as the vehicle approaches the right-lane boundary as shown in Figs. 2C and F. Therefore, the proposed LDI system focuses on measuring the orientation and position of lane boundaries to determine whether or not the LD occurs. Full-size image (27 K) Fig. 1. The basics of LDI: (A) description of the relationship of a CCD camera and lane, (B) orientations and positions of lane boundaries, and region of interest (ROI) for image processing, and (C) camera installation on a vehicle. Figure options Full-size image (35 K) Fig. 2. The phenomenon of LD: (A) LD of a vehicle to left side, (B, E) changes in θl and ρl when a vehicle approaches the left-lane boundary, (C, F) changes in θr and ρr when a vehicle approaches the right-lane boundary, (D) LD of a vehicle to right side. Figure options To realize the new LDI system, some constraints with respect to camera installation on a vehicle are required such that the optical axis of a camera mounted on a vehicle is as close as possible to the centerline of the car-body as illustrated in Fig. 1C, and as parallel as possible to the road surface. Such constraints are deeply related to setting of the ROI shown in Fig. 1B. Our previous edge distribution function (EDF)-based LDI system  and  was also based on such similar observations. However, in LDI, we only used the departure measure ξ defined by ξ = (θ1 − 90°)/(90° − θ2) where angles θ1 and θ2 are defined as shown in Fig. 1B. The departure measure ξ is equal to the ratio θl/θr. The EDF-based method has a weakness such that when a vehicle travels at the center of lane in a curved lane with broken lane marks, a big difference between the angles θl and θr occurs abruptly, the ratio θl/θr deviates in accordance from the value of 1, and eventually leads to a false alarm of LDI. The proposed method in this paper overcomes the weakness by introducing a linear regression (LR)  and reinforcing a new departure ratio, ρl/ρr. Different from the angles θl and θr, the positions ρl and ρr do not change abruptly in a non-departure situation in a curved road. Identifying the LD based on lane-related parameters is effective depending on image processing techniques even though their robustness is affected by noise factors. Much research about road-lane recognition by image processing has been conducted. From such research, feature-based methods , , ,  and , neural networks-based methods  and , and probabilistic methods  have been developed. Most of the research shares and combines various principles from these methods. Measuring the relative position between lane boundaries and a vehicle has been a well-known method for the design of LDI because of its simple concept , , ,  and . This method relies on the localization of lane marks to obtain the offset between the center of the lane and the central axis of a car body and often needs camera calibration and both widths of the vehicle and the lane. Hence, the method is affected by several parameters—the selection of a camera, lens optics, the mounting position of the camera, road types, and the subject vehicle. From this point of view, the RALPH  can be considered a similar approach to this method and consequently needs adjusting whenever the lane width is changed. Different from the extraction of positional offset between lane boundaries and a vehicle as done in RALPH  and Dickmanns and Zapp , the proposed approach is neither influenced by the width of the lane nor is it affected by vehicle-related data such as vehicle width and weight. Furthermore, curvature, time to lane crossing, coordinate transformation  and , and road model  and  are unnecessary in the proposed approach.
نتیجه گیری انگلیسی
There are two fundamental issues in judging whether or not the LDI system is successful: (1) to obtain reliable lane-related parameters in diverse conditions, and (2) to determine the beginning and ending of LD without a false alarm or miss-detection. From these aspects, the proposed LDI system was successful. Even though the LBPE was heuristic, it extracted candidate pixels to be on lane boundaries in a natural outdoor road scene robustly and sorted the extracted pixels into inside and outside pixels of a lane-mark, which, in the long run, supplied with sufficient data in deciding LD with robustness. The eight lane-related parameters overcame the weak points of the previous EDF-based LDI system. We minimized the erroneous results such as a false alarm or a missed detection with respect to LDI by introducing a multiple frames-based linear regression and fusing the regression results with the departure ratios in deciding the beginning and end of LD, and tracking the direction of departure. We showed the successful results of the proposed LDI system with some real images in various conditions.