TauAud: Test Augmentation of Image Recognition in Autonomous Driving

Deep learning has made great progress in autonomous driving. However, due to the insufficient and unbalanced data used for training and testing, deep learning model will easily overfit and have poor generalization capabilities. The problem of collecting balanced and sufficient data often arises because it is difficult and dangerous to collect data under abnormal situations, such as abnormal weather or camera damage. In addition, neural networks need to be trained on a large amount of accurate and reliable data. Incorrect data may result in poor performance. In this study, we analyze the impact of abnormal situations on the traffic images and perform a series of image processing methods on the normal traffic dataset to generate sufficient and balanced images. Moreover, we use the trained deep learning model to evaluate the effectiveness of the augmented dataset.