2D Line Filters for Vision-based Lane Detection and Tracking

2D Line Filters for Vision-based Lane Detection and Tracking
Anselm Haselhoff and Anton Kummert
Abstract— In this work a vision-based lane detection system is presented. The main contribution is the application of 2D line filters for lane detection which suppress noise in the near distance without destroying lines in the fare distance. Line pairs are decided to belong to lane markings by means of parallelism is world coordinates and reasonable constraints concerning the road width. The polygon constructed from the detected lane markings is then tracked over time via a kalman filter. Results of the approach are presented in terms of the detection accuracy on a labeled video sequence.
Faculty of Electrical, Information and Media Engineering, Chair of Communication Theory, University of Wuppertal, 42097 Wuppertal, Germany, haselhoff@uni-wuppertal.de, kummert@uni-wuppertal.de
˜ the 2D summation is Finally, for the two vectors r and r defined as
˜ ≤r r
˜ ) = ∑ ∑ f (x ˜, y ˜). ∑ f (r
x ˜=0 y ˜=0
x
y
(2)
III. L ANE D ETECTION S YSTEM In general the lane detection system is based on 2D line filters to extract lane features, the linear lane model presented in [4] and finally a kalman filter to track the lanes over time. To perform the detection a monochrome video camera with a resolution of 640 × 480 is used that is mounted at the position of the rear-view mirror. The whole lane detection system is visualized in figure 1. The components of the system are described in the follow subsections.
978-1-4244-2798-7/09/$25.00 ©2009 IEEE
0
m-1
M-1
x
0
x0 r1 A
x0+m-1 r2
x
y0
n-1 N-1 y
r rmax
y0+g. 2: Flat ground assumption.
B. 2D Line or Haar Filter The use of 2D line or Haar-like filters for object detection was first introduced by Viola & Jones [6]. The advantage of these features is a very fast computation due to the use of the integral image. One rectangular filter mask that is important for line detection is shown in figure 3. The filter output is calculated as a weighted sum of gray-level values contained in the black and white rectangles. The filter in Fig. 3 is related to the discrete filter approximation of a second order derivative and thus is denoted by vertical line filter hxx (r).
I. I NTRODUCTION Today’s vehicles contain driver assistance systems which observe the environment of the vehicle to provide road traffic with more comfort, safety, and efficiency. One example are lane detection and lane departure warning systems. Amongst others, video sensors are subject of current research in the lane detection context since video sensors enable sophisticated image processing algorithms and keep the costs to an affordable limit. The video camera facilitates application of various detection algorithms. The lane detection in [1] uses the sobel operator for edge detection and to handle fine and coarse structures they use an image pyramid. In [2] the edge detection is performed by the Canny filter and a fixed Gaussian filter is used for noise suppression. The work of [3] uses a Canny filter and in addition a Laplacian of Gaussian (LoG) edge filter. In best case these approaches handle noise by a fixed low-pass filter which is the same for high and low distances in world coordinates. The drawback is that fine structures in the fare distance can get lost. A good approach to handle the noise is presented in [4], where a stereo-camera is used in combination with distance dependent 1D edge filters. The lane detected is done by means of a Dark-LightDark (DLD) transition, where a gradient pair must have similar magnitude but an opposable sign. In our method we follow a similar strategy as in [4]. The advantage of our method is the usage of the integral image for image filtering, which can be used in other applications like vehicle [5] or pedestrian detection. The line filters enable that DLD transitions can be directly indentified by the sign of the filtered image with no additional computation. The remainder of the paper proceeds as follows. After a the definitions in section II the lane detection system, including the line filters, is described in section III. Finally, the results and conclusions are presented in section IV.
II. NOTATION AND BASICS In the following sections f (r) = f (x, y) denotes an M × N image with r = (x, y)T , defined for x = 0, 1, ..., M − 1 and y = 0, 1, ..., N − 1. Furthermore special unit vectors are defined as e1 = (1, 0)T , e2 = (0, 1)T , and e3 = (1, 1)T . ˜ = (x ˜, y ˜)T the lower equal For two vectors r = (x, y)T and r sign is defined as ˜≤r⇒x r ˜ ≤ x∧y ˜ ≤ y. (1)