Sensor-fusion and deep neural networks for autonomous UAV navigation within orchards

With the increase of population in the world, the demand for quality food is increasing too. In recent years, increasing demand and environmental factors have heavily influenced the agricultural production. Automation and robotics for fruit and vegetable production/monitoring have become the new standard. This paper discusses an autonomous Unmanned Aerial Vehicle (UAV) able to navigate through rows orchard rows. The UAV is comprised of a flight controller (AP stack), a microcontroller for analog reading of different sensors, and an On-Board Computer (OBC). Pictures are taken through a camera and streamed through WiFi to a Ground Control Computer (GCC) running a convolution neural network model. Based on prior training, the model outputs three directions: RIGHT, LEFT and STRAIGHT. A moving average of multiple frames per second is extracted and sent to a build-in Proportional-Integral-Derivative (PID) controller on the UAV. After error correction from this feedback, controller sends the direction to the flight controller using MAVLink protocol's radio channel overrides, thus performing autonomous navigation.