Real-Time Pothole Detection Using SwiftFormers

Potholes pose a significant challenge to global road infrastructure, impacting traffic flow and safety. Detecting these anomalies is crucial for accident prevention and damage mitigation. Despite previous efforts utilizing diverse imaging technologies like 2D and 3D methods, including CNNs, persistent challenges stem from limited data availability, the unpredictable nature of potholes, and delays in detection. Consequently, there’s a pressing need to develop more efficient and adaptable models that merge CNNs and transformers’ strengths, particularly for devices with limited computational resources. Our creation, using "SwiftFormer", integrates a novel efficient additive attention mechanism as its core, coupled with a path-aggregated FPN for pothole detection. In our dataset, this model achieved an mAP value of 74 and an AP50 value of 96.