We describe a framework for global shipping container monitoring using machine learning with low-power sensor hubs and infrared catadioptric imaging. A mesh radio satellite tag architecture provides connectivity anywhere in the world, with or without supporting infrastructure. We discuss the design and testing of a low-cost, long-wave infrared catadioptric imaging device and multi-sensor hub combination as an intelligent edge computing system that, when equipped with physics-based machine learning algorithms, can interpret the scene inside a shipping container to make efficient use of expensive communications bandwidth. The histogram of oriented gradients and T-channel (HOG+) feature is introduced for human detection with low-resolution infrared catadioptric images, and is shown to be effective for various mirror shapes designed to give wide volume coverage with controlled distortion.