ACF Based Feature Extraction and Mixture of Expert CNNs for Pedestrian Detection

Abstract

As the community strives towards developing systems with increasing levels of autonomy, the problem of detecting pedestrians in real world scenarios has become a pivotal problem to ensure safety concerns in domains like self driving cars and assistive robotics. While handcrafted features were the norm until a few years ago, deep learning has lead to an unprecedented surge in accuracies on benchmark datasets. This paper proposes a two stage pedestrian detector, wherein the first stage involves a cascade of Aggregated Channel Features (ACF) to extract candidate pedestrian windows from an image.  This is followed by introducing a thresholding technique on the ACF confidence scores that isolates candidate windows lying at the extremes of the ACF score distribution. The windows with ACF scores in between the upper and lower bounds are passed on to a Mixture of Expert (MoE) CNNs for more refined classification in the second stage. Results show that the designed detector yields close to  state-of-the-art performance on the INRIA and CALTECH benchmark datasets[3] and yields a miss rate of 10.35% and 23.6% respectively at FPPI=10−1 .