A context-dependent vision system for pedestrian detection

Abstract

Robustness is a key issue in pedestrian detection for autonomous vehicles. Contextual information, if well exploited, should increase robustness and performance. Specifically, contextual knowledge allows for the integration of algorithms performing well only in specific situations, which would otherwise be excluded from a system designed for the general case. Here, we discuss using context in a vision-based system. Contextual evolution of scene parameters is represented as the hidden process of a Hidden Markov Model. Consequently, a Bayesian framework is adopted for all principal elements, including sensor models for specialised algorithms and sensors observing the current context. Our strategy allows re-use of known algorithms, at the same time enabling context-sensitive developments.