Visual Tracking Based Deep Learning and Control Design Onboard Small-Sized Quadrotor UAV
Kuoh Patrice Gama, Chaofang Hu, Hao Wu · 2022 41st Chinese Control Conference (CCC) · 2022
It is known that unmanned aerial vehicle (UAV) with computer vision based deep neural network (DNN) capabilities, are extremely promising for applications such as surveillance, path planning and navigation. Unfortunately, DNN's computational and power requirements are still higher than the budget for small-sized UAV due to onboard memory constraints. In this paper, we present the design and implementation of a visual control steering in closed-loop end-to-end target tracking based DNN and directly utilizing the onboard hardware resources of the small-sized quadrotor UAV. This is archived, by adopting convolutional neural network (CNN) models such as, MobileNet combine with single shot multibox detection (SSD) algorithm to solve the problem of real-time object tracking onboard. The network is compressed (optimized) to run in a low powered ARM architecture by employing quantization and cortex microcontroller software interface standard (CMSIS-NN). Finally, a color feature extraction (blob-like-features) is introduced to improve tracking and control accuracy. All the image processing is done real-time onboard, the flexible navigation system can span a wide performance range, making it a feasible solution for real-time tracking-control onboard a small-sized UAV.