Safety-Aware Metric for People Detection

Abstract

At construction sites and off-road scenarios people are often in danger, because drivers of heavy machinery have to focus on their tasks and might recognize dangerous situations too late. They, therefore, heavily rely on driver assistance systems. For people detection in such environments not all inference results (true positives, false positives or false negatives) are equally relevant. In this paper we propose a novel metric for people detection, providing an increased emphasis on human hazard prevention. Most works blindly allocate efforts for improving detection performance using situation-agnostic metrics like mean average precision or F1-score. The objective of this work is to determine solid pragmatic variables that directly relate to a person's safety and judiciously combine them, coming up with a safety-aware metric for people detection. To set up a system focusing on safety and reliability, the results of a neural network have to be subsequently verified. We therefore propose a new detection plausibility check based on an extension of the stixel algorithm and a multi-object tracker. Our results are evaluated on a new self-recorded off-road dataset for people detection and compared with standard evaluation metrics. Furthermore, results obtained from training are compared against the proposed plausibility algorithm.