Lightweight Real-time Detection of Components via a Micro Aerial Vehicle with Domain Randomization Towards Structural Health Monitoring

Abstract

Civil structural component detection plays an integral role in Structural Health Monitoring (SHM) pre and post-construction. Challenges including but not limited to labor-intensiveness, cost, and time constraints associated with traditional methods make it a less optimal approach in SHM. Despite the success of deep convolutional neural networks in diverse detection problems, the required computational resources are a challenge. This has led to rendering a chunk of resource-constrained edge nodes less applicable with deep convolutional neural networks. In this paper, a computational-efficient deep convolutional neural network is presented based on Gabor filters and a color Canny edge detector. Generic Gabor filters are generated and used as initializers in the computational-efficient deep convolutional neural network presented, afterward trained on building components data. Next, extensive offline and online experimentation with a resource-constrained edge node is conducted and evaluated using diverse metrics. The computational-efficient detection model demonstrates to be effective in detection and via NVIDIA GPU profiler, we observe conservation of around 30% of computational resources during training. The computational-efficient detection model adduces almost a 3% mean average precision higher than two state-of-the-art detectors and records a promising frame processing rate during the online experimentation.