Back
to indexMatthies, L. and Grandjean, P.
Summary
This paper presents a quantitative methodology for measuring perception accuracy with stereo vision, or it can be applied to LADAR sensors. It strives to quantitatively answer 2 questions: how good is the range data itself and how good are the obstacle detection algorithms applied to the data. Using these methods, a perception mechanism can be accurately measured for error in the preliminary design stage. This paper is widely cited and provides a mechanism to evaluate different sensors and algorithms up front, before development begins.
Methods
The basic question to answer are how far ahead do we have to detect obstacles, and what resolution is needed to detect obstacles at that distance? The minimum look-ahead distance is calculated as:
dl = dc + v0 (2tc + ta) + (v0 * v0)/(2a)
Where v0 is initial velocity, a is deceleration ration, tc and ta are latency times for perception and actuation, respectively, and dc is the total distance from the nose of the vehicles to the camera. This is a great formula to tell us how far ahead we have to look.
The rest of the paper calculates the quality of the range data and the quality of the obstacle detection algorithms. Without going into the math here, the general error rate is calculated per pixel, and reduced by increasing the resolution of the image. Statistically there are several assumptions made, such as all obstacles are step functions, which may not be valid for our circumstances. Mathematically, measurements are taken of known distances to objects and the error rate is calculated to generate quality of range data and obstacle detection.
Paper is written very technically with little use of clear english. Its a good metric, but difficult to understand and parse. Recommend math folks with strong probability background take a swing at understanding the details presented here.
Keywords
obstacle detection, range data, error rate calculation, performance model, quantitative analysis of quality
Rating
6
Bibtex Entry
@article{ matthiesstochastic,
author = "Matthies, L. and Grandjean, P.",
title = "Stochastic Performance Modeling And Evaluation Of Obstacle Detectability With Imaging Range Sensors",
journal = "T-RA",
volume = "10",
pages = "783-792",
url = "citeseer.nj.nec.com/matthies94stochastic.html"
}