Road region segmentation based on sequential Monte-Carlo estimation

Abstract

Road region is an important information for guidance of autonomous vehicles or robots. The goal of this research is to develop robust monocular algorithm suitable for region based road detection. This paper deals with a problem, how to estimate probability density function (pdf) of road region appearing in sequential images. The key idea is to construct pdf from temporal sequence of observations throughout the image sequence, where pdf has color components and spatial coordinates as its variables. The problem of pdf estimation was formulated in terms of Bayesian filtering, and sequential Monte-Carlo method was addopted as a tool to solve the problem. The proposed method was evaluated on real image sequences, and effectiveness of the proposed method is demonstrated.

Road Tracking Method Suitable for Both Unstructured and Structured Roads

“…As was mentioned in section 3.3, we used same segmentation parameters for all sequences. Consistently with our previous results [16], the proposed segmentation method works well with one set of parameters for various types of roads. Opposite to this, the region growing segmentation used in REF-1 is very sensitive to a proper threshold of the pixel aggregation criteria, and we could not find a common compromise value, giving us good segmentation results for all sequences.…”

Pathway estimation for vision based road following suitable for unstructured roads

“…As a part of our previous work, we have developed a road region segmentation method based on sequential Monte-Carlo estimation [14]. A road region in this method is expressed in a form of a probability density function (pdf) of coordinates and color components.…”

Sequential Monte-Carlo Based Road Region Segmentation Algorithm with Uniform Spatial Sampling

“…To build-up a classifier for decision between road and non-road pixels in the captured image, we have to handle such changes in an adaptive way. A method based on recursive estimation of a statistical distribution of the road features has been already developed in the earlier work [13]. Although this method can potentially avoid some issues discussed in the Section 1, and it is quite robust to parameter tuning, it still doesn't produce satisfactory results under certain circumstances.…”

“…As an approach to this issue, a road region segmentation method based on sequential Monte-Carlo (SMC) has been developed in the previous work [13]. With this method, the estimation of the pdf of road features can be performed in an adaptive way, and it also introduces a systematic solution to the above mentioned "chicken-and-egg" problem.…”