Autonomous Road Defects Segmentation Using Transformer-Based Deep Learning Models With Custom Dataset

Potholes and cracks on road surfaces pose significant safety hazards and lead to increased vehicle maintenance costs and reduced roadway lifespans if not promptly addressed. Automated detection and segmentation of these road defects are essential for efficient maintenance planning and hazard mitigation. In this paper, we present a comprehensive evaluation of transformer based segmentation models applied for the detection and segmentation of potholes and cracks on different road surfaces. We introduce a custom dataset specifically modeled for urban roads that can be used for low-middle income countries with poor infrastructure and low maintenance resources. It features diverse road images under varying lighting, weather, and road conditions. We compare the performance of various segmentation models in terms of segmentation accuracy and robustness on this dataset. Our experimental results reveal significant performance variations across model architectures: larger models (M12) achieved the highest precision of 0.621 for bounding boxes and 0.618 for masks, while smaller models (M1) demonstrated computational efficiency with precision scores of 0.491 and 0.428 respectively. The mAP@50 scores ranged from 0.225 to 0.314 across all models, with Family III variants showing mixed performance profiles. Computational benchmarking revealed inference speeds ranging from 25-35 FPS for large models to 85-120 FPS for lightweight variants on NVIDIA RTX A6000 hardware. The results highlight the strengths and limitations of different models, providing insights for researchers and practitioners in the field of road maintenance and safety.