A tool for vision based pedestrian detection performance evaluation

Abstract

This paper describes a system for evaluating pedestrian detection algorithm results. The developed tool allows a human operator to annotate on a file all pedestrians in a previously acquired video sequence. A similar file is produced by the algorithm being tested using the same annotation engine. A matching rule has been established to validate the association between items of the two files. For each frame a statistical analyzer extracts the number of mis-detections, both positive and negative, and correct detections. Using these data, statistics about the algorithm behavior are computed with the aim of tuning parameters and pointing out recognition weaknesses in particular situations.