Capturing Object Detection Uncertainty in Multi-Layer Grid Maps

Abstract

We propose a deep convolutional object detector for automated driving applications that also estimates classification, pose and shape uncertainty of each detected object. The input consists of a multi-layer grid map which is well-suited for sensor fusion, free-space estimation and machine learning. Based on the estimated pose and shape uncertainty we approximate object hulls with bounded collision probability which we find helpful for subsequent trajectory planning tasks. We train our models based on the KITTI object detection data set. In a quantitative and qualitative evaluation some models show a similar performance and superior robustness compared to previously developed object detectors. However, our evaluation also points to undesired data set properties which should be addressed when training datadriven models or creating new data sets.− log E(ŝ) ≈ 1 J J j=1

Uncertainty Estimation in One-Stage Object Detection

“…One model with dropout enabled (as described above), one without dropout but the aleatoric loss enabled and one combined model with dropout and aleatoric loss enabled (referred to as aleatoric+epistemic). The initial training without dropout and aleatoric loss is crucial for stable training as found by [4] and confirmed by us.…”

“…Wirges et al [4] also employed a two-stage approach but also added dropout to some of the convolutional layers. When applying dropout to all layers, they found that the model no longer converges.…”

“…For our experiments we use the EuroCity Persons (ECP) dataset [15] which consists of roughly 47000 images of which around 7000 are captured at night. In total around 218000 pedestrians and 20000 riders are labeled 4 . The dataset provides annotations for the level of occlusion of every object.…”

“…Recently their methods were incorporated into object detection by [3], [4]. Both used two-stage models which first generate proposals from a region proposal network and, in a second network, refine the box offsets and predict a class.…”

MonoLoco: Monocular 3D Pedestrian Localization and Uncertainty Estimation

“…Compared to previous methods [25,53], we design a Laplace loss which works with relative distances to keep into account the role of distance in our predictions. Estimating the distance of a pedestrian with an absolute error can lead to a fatal accident if the person is very close, or be negligible if the same human is far away from the camera.…”