292 
A TIME-CONSTRAINED VLSI STEREOVISION SYSTEM 
Yang NI*, Bogdan ARION*, Francis DEVOS^ 
"Institut National des Télécommunications, EPH-CAO, 91011 Evry, FRANCE 
ni@galaxie.int-evry.fr ; arion@galaxie.int-evry.fr 
+Université Paris Sud, 91405 Orsay, FRANCE 
devos@ief-paris-sud.fr 
KEYWORDS: stereovision, real-time, parallel processing 
ABSTRACT 
This paper presents an analog VLSI implementation of a time-constrained stereovision algorithm suited to obstacle 
detection applications for autonomous robots or self-guided vehicles. 
The algorithm is based on spatio-frequency analysis of stereopair images, using local extremum extraction from the 
bandpass filtered images as matching primitives, instead of zero-crossing points. Although this algorithm could be 
implemented on classical CCD-DSP based vision systems, its simplicity and regularity on one hand and the real-time 
context of the application on the other hand motivate an analog VLSI retina-like solution. 
Therefore we have designed a completely new line-parallel processing interfaced with an on-chip photoreceptor 
matrix. The two key elements in this analog implementation are the original DoG analog filter and the local extremum 
extraction circuit. This stereo retina has been simulated with real-scenes images taken from a running car on a 
highway. The results are very positive and a 128-pixel line retina has been designed and submitted to fabrication. 
1. INTRODUCTION 
Although binocular stereoscopy is known to give reliable information for a 3-D reconstruction from two 2-D images, 
the impact on machine based vehicle guidance is limited by the amount of computation in most of the present 
algorithms or by the high cost of the hardware implementations. 
The field of stereovision is extremely large and a global solution is hard to find. It is therefore admitted that different 
applications need different results from the stereo module. Consequently the computation complexity is not the same, 
and we can identify two major stereovision depending tasks. ; 
a) the recognition task involves an important amount of computation in order to reconstruct the visual surface. 
In this context it is a time-relaxed task. 
b) the navigation task focuses on obstacle detection in real-time environments, therefore the time constraint 
becomes essential. From this point of view, time-constrained applications admits more coarse-grained structures on 
the resulting depth map. 
Our application drop in the second case, as long as avoiding collisions is more important than knowing the real 
nature of the potentially dangerous obstacles. 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommision Workshop "From Pixels to Sequences", Zurich, March 22-24, 1995