Information Technology Reference
In-Depth Information
This paper is organized as follows: subsequent to the introduction, a brief
literature review is conducted citing some of the current lane detection imple-
mentations. Then, the different components used in the detection and track-
ing of lane markers are explained. Finally, the method for calculating errors is
described. The performance of the proposed system is assessed on real world
videos recorded at various times of the day. Finally, the conclusion and planned
improvements are discussed.
2PorRch
Lane detection is still an active area of automotive research. Conventional ap-
proaches suggest the application of thresholds after studying patterns in his-
tograms in hopes of segmenting lane marker pixels from background or road
pixels [1,2]. Unfortunately, histogram approaches are vulnerable to outlier inten-
sity spikes. The use of edge images to find lines or curves using a variety of kernel
operators has been suggested by [1,3,4,5,6] but face diculty when markers show
signs of age and wear. A piece-wise Hough transform to fit a line on a curve has
been used to handle conditions involving scattered shadows [7,8]. Additionally,
the incorporation of edge directions has been used to remove some false signalling
[2,6]. Unfortunately, invariance to scale and rotation tends to be major problem
for these methods. Classifying small image blocks as lane markers using learning
methods has been suggested by [9]. But a good quality linear classifier is di-
cult to derive without an infinitely large catalog of negative training examples.
Lane detection using adaptive thresholds and one dimensional iterated matched
filters has been suggested by [10,11]. Unfortunately, one dimensional template
matching did not perform so well during the day.
Lane detection is a crucial component of many DA systems; thus, it needs
to be extremely reliable and robust. Current research appears to boast high
performance only in the presence of favorable illumination and road surface
conditions. Unfortunately, these conditions are unlikely to exist on the road
network in most big cities. Based on the literature survey, it can be seen the
feature extraction stages in existing implementations are unable to satisfactorily
discriminate between lane markers and surface artifacts. Consequently, there is
a need to develop improved techniques to detect lane markers that is able to
cope with the variety of road conditions that exist around the world.
3SyemOrvew
The overview of the proposed lane detector is shown in Fig. 1. First, the images
undergo preprocessing the form of Inverse Perspective Mapping (IPM) and color
conversion. Then, template matching in addition to morphology and ellipse pro-
jection finds areas containing lane markers. Finally, the Kalman filter is used to
track lane marker estimates while RANSAC helps eliminate outliers.
Search WWH ::
Custom Search