Chunks: The Remedy for Notorious False Alarms in Pedestrain Detection

Abstract

Though significant progress has been made in last decade but still pedestrian detection is a challenging task. In real world, pedestrians are bound to produce artifacts, like pose & attire variations and occlusions, which are some of the main causes of false alarms. We propose a method which can tackle these variations efficiently. Instead of traditional deformable parts model, we propose random patches (called chunks) to capture the features properties of pedestrians. We have cascaded chunks with Aggregate Channel Features (ACF) detector in order to ratify the pedestrian hypothesis generated by ACF. Our method gives the miss rate of 16.51% at 10-1 false positives per image under reasonable condition, which is among one of the best results achieved on INRIA pedestrian dataset. Our method also improved on the same dataset under partial and heavy occlusion condition.