Validation of Pedestrian Detectors by Classification of Visual Detection Impairing Factors

Abstract

Validation of AI based perception functions is a key cornerstone of safe automated driving. Building on the use of richly annotated synthetic data, a novel pedestrian detector validation approach is presented, enabling the detection of training data biases, like missing poses, ethnicities, geolocations, gender or age. We define a range of visual impairment factors, e.g. occlusion or contrast, which are deemed to be influential on the detection of a pedestrian object. A classifier is trained to distinguish a pedestrian object only by these visual detection impairment factors which enables to find pedestrians that should be detectable but are missed by the detector under test due to underlying training data biases. Experiments demonstrate that our method detects pose, ethnicity and geolocation data biases on the CityPersons and the EuroCity Persons datasets. Further, we evaluate the overall influence of these impairment factors on the detection performance of a pedestrian detector.