An evidential framework for pedestrian detection in high-density crowds

Abstract

This paper addresses the problem of pedestrian detection in high-density crowd images, characterized by strong homogeneity and clutter. We propose an evidential fusion algorithm which is able to exploit multiple detectors based on different gradient, texture and orientation descriptors. The evidential framework allows us to model the spatial imprecision arising from each of the detectors. A first result of our study is that the fusion results underline clearly the good complementarity among the four descriptors we considered for this specific context. Moreover, the proposed algorithm outperforms a fusion solution based on Multiple Kernel Learning on difficult high-density crowd images acquired at Makkah at the height of the Muslim pilgrimage.

Belief Function Definition for Ensemble Methods - Application to Pedestrian Detection in Dense Crowds

“…In the context of high-density crowd pedestrian detection, in [3] we propose a robust fusion strategy also based on the belief functions framework, that is able to take into account the spatial imprecision of each different classifier, and we show the improved performance of this method with respect to MKL and the straightforward product of probabilities. Spatial imprecision is introduced following the work of [21], that allows to define a valid bba using morphological operators, exploiting the duality property between erosion and dilation (respectively opening and closing) and Bel and P l values for the final bba.…”

“…Popular particularly in the field of stereo matching, the DAISY [8] descriptor has been successfully employed for the first time for head detection in difficult crowd scenes in [3].…”

“…Among some descriptors which are widely used in pedestrian detection, those which are the best suited for detection in high-density crowds have been recently highlighted in [3].…”

“…As in [3], a spatial Gaussian structuring element fitted in a window of radius a is considered, to better take into account the spatial consistency.…”

“…For this reason, following the preliminary work we introduced in [3], we use bbas to model the spatial imprecision due to the close resolutions of object (head) and descriptor respectively. Applied in the spatial domain, the work of [21] allows indeed to associate to every pixel p of the image a bba m Now, to introduce spatial imprecision, we derive bbas performing erosion and dilation (respectively opening and closing), but this time in the image domain, that is to say taking into account neighbor pixels based on the assumption that they are likely to belong to the same class.…”