Neural network approach for vision-based track navigation using low-powered computers on MAVs

A quadrotor Micro Aerial Vehicle (MAV) is designed to navigate a track using neural network approach to identify the direction of the path from a stream of monocular images received from a downward-facing camera mounted on the vehicle. Current autonomous MAVs mainly employ computer vision techniques based on image processing and feature tracking for vision-based navigation tasks. It requires expensive onboard computation and can create latency in the real-time system when working with low-powered computers. By using a supervised image classifier, we shift the costly computational task of training a neural network to classify the direction of the track to an off-board computer. We make use of the learned weights obtained after training to perform simple mathematical operations to predict the class of the image on the onboard computer. We compare the computer vision based tracking approach with the proposed approach to navigate a track using a quadrotor and show that the processing rates of the latter is faster. This allows low-cost, low-powered computers such as the Raspberry Pi to be used efficiently as onboard companion computers for flying vision-based autonomous missions with MAVs.