A lazy decision approach based on ternary thresholding for robust target object detection

One of the main problems of binary classification of overlapping distributions is that there always exist misclassification errors with any value of threshold. In this paper, we propose a novel lazy decision approach for robust object detection and tracking, where decision on an uncertain observation whose evaluation lies between low and high thresholds is postponed until a clear evidence appears. As a practical application of the proposed approach, we present a sensor fusion pedestrian detection system for safe navigation of UGVs in driving environment. We combine a laser-based detection of target candidates and vision-based evaluation within the proposed lazy decision framework. Experimental results on real test data demonstrate effectiveness of the proposed approach, showing significant improvement of precision-recall performance.