Metamorphic Testing for the Validation of Deep-learning-based Object Detection Systems

Complex tasks, such as environment perception in autonomous driving, are increasingly being taken over by systems based on deep neural networks.Testing such systems is challenging and requires various test methods to ensure sufficient validation in series development.IAV and TU Munich have applied Metamorphic Testing, a validation method that does not require ground-truth information, to lidar systems and tested this for use in object detection functions.g In autonomous driving, Deep Neural Networks (DNNs) are increasingly being used for object detection.Systems based on DNNs are often considered as black boxes because the patterns and decisions learned by the system are not transparent to humans.A huge challenge for developers trying to thoroughly test DNN-based object detection, is the vast number of complex scenarios and traffic conditions.Additionally, for many requirements to be testable, a test oracle must assess the correctness of the test result.For example, high-precision reference sensors for detecting the environment are used to generate a test oracle for the validation of object detection algorithms.