Embedded port infrastructure inspection using artificial intelligence

This paper is related to the H2020 RAPID project, focusing on the AI automated monitoring of critical port infrastructure such as concrete structure. An important objective in RAPID was to translate a technical expertise of labelling cracks into a UAV real-time embedded solution based on deep neural networks. The efficiency of a deep learning algorithm is highly dependent on the data used for training, and this paper illustrates the fact that the use of open-source data is not sufficient. An intensive collaboration between neural network and industry experts made it possible to obtain a relevant data set of sufficient size to carry out quality training. This collaborative work also allowed the definition of ground truths, necessary for the validation of the detection system. In this paper, we provide a definition of the useful metrics and objectives for the algorithms in accordance with the complexity of the cracks and their environment, used to identify the best neural network in terms of efficiency, and performance to embed it on a UAV. Our research then focused on the hardware platform that could be used as an onboard computer for the drone, considering Size, Weight and Power (SWaP) constraints. We applied optimization methods to reduce the latency of our models while maintaining high accuracy. These techniques allowed us achieve a state-of-the-art detection rate while complying with the real-time requirements of the overall system, and the need to increase productivity of mission inspections in a port environment through high-speed inferences.