People in the weeds: Pedestrian detection goes off-road

Abstract

Robotics offers a great opportunity to improve efficiency while also improving safety, but reliable detection of humans in off-road environments remains a key challenge. We present a person detector evaluation on a dataset collected from an autonomous tractor in an off-road environment representing challenging conditions with significant occlusion from weeds and branches as well as non-standing poses. We apply three image-only algorithms from urban pedestrian detection to better understand how well these approaches work in this domain. We evaluate the Aggregate Channel Features (ACF) and Deformable Parts Model (DPM) algorithms from the literature, as well as our own implementation of a Convolutional Neural Network (CNN). We show that the traditional performance metric used in the pedestrian detection literature is extremely sensitive to parameterization. When applied in domains like this one, where localization is challenging due to high background texture and occlusion, the choice of overlap threshold strongly affects measured performance. Using a permissive overlap threshold, we found that ACF, DPM, and CNN perform similarly overall in this domain, although they each have different failure modes.