On Diversity and Complementarity of Pedestrian Detection Models

Abstract

Pedestrian Detection (PD) is now almost a mature field of signal processing in which numerous application areas exists. Emerging one of these applications is driving autonomy that requires advanced driver assistance systems (ADAS) for PD. Although various benchmark data sets are collected and several methods are already developed, the reported results show that current systems lack generalization capacity due to strong dependence of system performance on PD training set. In other words, the systems trained with a pedestrian data set usually perform considerably well on a test set taken from the same benchmark, but their performance drop drastically when tested on a different benchmark data set. In order to overcome this limitation, one of the approaches might be fusing outcomes of different systems. Here, “different systems” can refer either to train the same model with different data sets or using entirely different models. In both cases, the outcomes of different systems should be diverse enough to cover the entire solution space. Up to our knowledge, the performance of existing well-established models for PD is frequently compared to each other but their diversity is not analysed in detail yet. In this study, the feasibility of PD systems, whose effectiveness on PD is shown by earlier studies, was examined by systematically training and testing them with different benchmark data sets and their combinations. For this purpose, pairwise and non-pairwise measures of diversity have been employed and several conclusions are driven about the complementarity of existing PD models.