Publications / 2024 Proceedings of the 41st ISARC, Lille, France
Pavement crack tracking in unstructured road environments stands as a crucial and ongoing challenge, playing a pivotal role in ensuring precise crack sealing for automated pavement repair. However, slender cracks often present issues with insufficient feature extraction and reduced tracking efficiency. This article proposes a dynamic visual tracking method for pavement cracks utilizing a hybrid adaptive control scheme combined with a self-tuning neural network and proportional-integral-derivative (PID). Specifically, the system capitalizes on a transformer-based crack segmentation system to extract crack features from the road image plane and determine an optimized control input to direct the robot. Segmentation module cross-integrates coarse-grained and fine-grained features using the self-attention (SA) and cross-attention (CA) for fusion operations within the fusion transformer module. Crack tracking is additionally equipped with a NeuralPID controller for adaptable control parameter adjustment. The effectiveness of this proposed method, tested thoroughly on a physical robot platform, illustrates significantly positive results in achieving real-time tracking of pavement cracks.