Low-Power Drone-Mountable Real-Time Artificial Intelligence Framework for Road Asset Classification

With the continuous increase in interstate highway traffic and demand for higher safety standards, there is a growing need for rapidly scalable road inspection. Currently, inspection and condition assessment of roadways involve manual operations which increase labor costs and limit the scalability and inspection coverage. Furthermore, manually inspecting highways adds additional safety risks for highway workers and road inspectors. To address these challenges, we envision a fully automated process of highway inspection. This paper presents a novel low-power drone-mountable real-time artificial intelligence (AI) framework for road asset classification through visual sensing, which is the first step toward a fully automated inspection system. We analyzed a state DOT dataset, consisting of 14 different kinds of defected road assets. To this end, we developed our baseline framework using MobileNet-V2, which is a convolutional neural network (CNN) specially developed for mobile and embedded platforms. Since our target dataset was small and CNNs networks require a huge amount of data, we leveraged transfer learning, by pretraining MobileNet-V2 using the ImageNet dataset and then fine-tuned it on our target dataset. This new framework was ported to embedded platforms Nvidia Jetson Nano with the capability to perform on-board drone processing. Overall, our results demonstrate 81.33% accuracy on the test set while processing 7.4 frames per second and occupying a total power of 1.9 W. It achieved a Power Reduction Factor (PRF) of 21.17 over Nvidia TitanV implementation, with only 8.74% impact on the projected drone flight time.