Deploying Edge AI for Road Surface Damage Detection Using TensorFlow Lite and Teachable Machine

Abstrak

Road damage compromises transportation safety and drives high infrastructure maintenance costs. To address the limitations of traditional methods, which are expensive and non-scalable, this study proposes an Edge AI alternative utilizing widely available smartphones and machine learning capabilities. We present a real-time road damage detection system powered by TensorFlow Lite and Teachable Machine. The system architecture employs lightweight CNN models (such as MobileNet and EfficientNet Lite) optimized for edge deployment. This implementation enables the immediate detection of anomalies (cracks and potholes) directly on the mobile device without cloud dependency, ensuring low latency. Testing demonstrated robust model performance. For pothole detection, the system achieved an accuracy of 95%; for cracks, the accuracy was 92%. During real-time trials in daytime urban settings, the system achieved an average detection latency of 200 milliseconds with an accuracy of 94%. This user-friendly system also supports data collection via crowdsourcing, facilitating comprehensive infrastructure monitoring and proactive maintenance. This innovation offers a scalable, cost-effective, and user-friendly solution with significant potential to advance transportation safety and maintenance efficiency.