Road traffic accidents have been always one of the key concerns in autonomous driving. However, there is a relative lack of research in the detection of road surface objects, particularly speed breakers, in the context of few-shot learning (FSL). Thus, this work proposed a novel framework by using Siamese networks combined with different pretrained networks, to tackle the aforementioned problem in an FSL regime. A set of 2207 images were collected using a dashcam and smartphone, belonging to several key classes, namely, speed bumps, rumble strips, painted arrows, zebra crossings, and miscellaneous objects, and from this set, several small datasets were created for experimentation purposes. Additionally, the dashcam view images were taken using the mounted camera positioned from the view height of a car driver. Our results demonstrate that a simple Siamese architecture can robustly tackle the detection and differentiation of our chosen road targets in an FSL regime when supported by powerful pretrained networks. The study provides a systematic comparison of the impact of different pretrained networks, showing general robustness across pretrained architectures, with different conditions showing a relative preference for specific architectures. For example, ResNet101 and DenseNet121 were shown to be particularly strong in the smallest FSL condition, with 300 training instances, respectively obtaining 97.15% and 96.97% average accuracy in a three-way one-shot classification task. Overall, this work provides supporting evidence and complements the usefulness of Siamese networks when combined with pretrained networks in the FSL regime for speed breaker recognition.
Few-Shot Learning Based Speed Breaker Recognition using Siamese Networks
Int. J. ITS Res.
International Journal of Intelligent Transportation Systems Research ; 23 , 2 ; 1218-1241
2025-08-01
24 pages
Article (Journal)
Electronic Resource
English
Rumble Strips , Few-shot learning , Siamese network , Speed bump Psychology , Automotive Engineering , Civil Engineering , Electrical Engineering , Psychology and Cognitive Sciences , Engineering , User Interfaces and Human Computer Interaction , Robotics and Automation , Computer Imaging, Vision, Pattern Recognition and Graphics
Few-Shot Learning Based Speed Breaker Recognition using Siamese Networks
Springer Verlag | 2025
|Deep Learning-Based Speed Breaker Detection
Springer Verlag | 2024
|SIAMESE-RESNET: IMPLEMENTING LOOP CLOSURE DETECTION BASED ON SIAMESE NETWORK
British Library Conference Proceedings | 2018
|