Tracking the pose of a swarm of insects in free flight is an important challenge necessary both for behavioural studies at scale, and also for testing interventions for the management of species harmful to human and animal health, including mosquitoes. Multi-camera techniques such as convex hull reconstruction have reconstructed pose of the limbs, head, body, wings and other parts of the animal in a lab setting successfully, but tracking in the field remains an important and difficult open question. Here we will demonstrate a promising new approach for background subtraction and state estimation of a mosquito in free flight in a variety of conditions. We use both camera recordings and synthetic ray-traced data to train a convolutional neural network to recognise a variety of backgrounds and estimate the state of a mosquito in free flight, showing robust performance with a variety of natural and artificial backgrounds.