SVM-based obstacle classification in visible and infrared images

This paper describes a Support Vector Machine (SVM)-based obstacle recognition system that can recognize both vehicles and pedestrians using bimodal vision. Different techniques were investigated in order to recognize the detected obstacles by the extraction of a compact and pertinent numeric signature from visible and infrared spectrum. A bi-objective optimization (using error classification rate and classification time) is employed to assure the SVM-parameters selection. A comparative study of individual visual obstacle recognizers versus fusion-based (at the feature, kernel and matching-score level) systems is performed. An important advantage of the fusion-based systems is their possibility to adapt to the environmental conditions due to a weighting parameter which establishes the importance of each sensor in a specific situation.