40 / 2021-12-06 17:49:00

Robustness Evaluation of Traffic Light Detection Models using Metamorphic Testing

Metamorphic Testing; Traffic Light Detection Models; Robustness Evaluation

Nowadays, the advent of driverless cars reduces the energy that people spend on driving. However, the safety of driverless vehicles has become a very important topic. Traffic light detection as an important part of Autonomous Driving System, its robustness is particularly important. In recent years, people have proposed many models and methods for Traffic Light Detection, most of which are based on convolutional neural network, but few people have evaluated the robustness of the model. Nowadays, there are many methods for robustness evaluation, but since the output of the neural network is unknowable, that is, there is a test oracle problem. Since the Metamorphic Testing can alleviate the test prediction problem, this paper adopts the Metamorphic Testing(MT) to evaluate the robustness of the Traffic Light Detection Models(TLDM). According to the characteristics of traffic lights and the actual scene of traffic lights, this paper puts forward three Metamorphic Relations(MR), and uses them to evaluate the most commonly used TLDM. The evaluate result is that the TLDM is not robust.