Abstract

Modern preview control-based suspension and advanced driver assistance systems increasingly rely on accurate road disturbance classification to anticipate hazards and optimize vehicle responses. This paper studies the application of deep learning models for classifying road anomalies. To this end, three state-of-the-art pre-trained deep learning architectures: MobileNetV2, ViT, and InceptionV3 are applied for the road anomaly classification task. The effectiveness of the three models is studied on a novel dataset developed as part of this study. The dataset comprises seven distinct categories of road disturbances typically encountered on urban and suburban roads in the Indian subcontinent. ViT achieves testing accuracy of 97.3% and Precision, Recall, and F1 scores of 0.97, demonstrating superior classification capabilities vis-à-vis the other models. MobileNetV2 achieves an accuracy score of 97.33%, but with relatively higher misclassification rates. InceptionV3 also exhibits robust performance, balancing accuracy and generalizability. This study demonstrates the superiority of ViT over InceptionV2 and MobileNetV3 for this problem, thereby highlighting the potential of vision transformer-based learning for road anomaly classification. The novel road disturbance dataset developed in this work would contribute to further related research in intelligent transportation systems. Furthermore, all code and datasets developed as part of this study are made openly available to support further research in this and related domains.

Authors

A.V. Satyanarayana, Anjnay Mahajan, Prem Prakash Vuppuluri, C. Patvardhan
Dayalbagh Educational Institute, India

Keywords

Deep Learning, MobileNetV2, ViT, InceptionV3, Preview Control, Road Anomaly