Deep learning-based UAV

Deep learning is a subfield of machine learning based on a set of algorithms that attempt to understand data like human being. In our research, we apply deep learning to solve different UAV problems, such as end-to-end perception, navigation, and control.

 

 

Neural Network MPC for quadrotor (NNMPC)

A dynamic model that considers both linear and complex nonlinear effects extensively benefits the model-based controller development. However, predicting a detailed aerodynamic model with good accuracy for unmanned aerial vehicles (UAVs) is challenging due to their irregular shape and low Reynolds number behavior. This work proposes an approach to model the full translational dynamics of a quadrotor UAV by a feedforward neural network, which is adopted as the prediction model in a model predictive controller (MPC) for precise position control. The raw flight data are collected by tracking various pre-designed trajectories with PX4 autopilot. The neural network model is trained to predict the linear accelerations from the flight log. The neural network-based model predictive controller is then implemented with the automatic control and dynamic optimization toolkit (ACADO) to achieve real-time online optimization. Software in the loop (SITL) simulation and indoor flight experiments are conducted to verify the controller performance. The results indicate that the proposed controller leads to a 40% reduction in the average trajectory tracking error compared to the traditional PID controller.

Youtube PDF

 

Autonomous Object Tracking UAV system (AUTO)

 

The ever-burgeoning growth of autonomous unmanned aerial vehicles (UAVs) has demonstrated a promising platform for utilization in real-world applications. In particular, UAV equipped with a vision system could be leveraged for surveillance applications. This paper proposes a learning-based UAV system for achieving autonomous surveillance, in which the UAV can be of assistance in autonomously detecting, tracking, and following a target object without human intervention. Specifically, we adopted the YOLOv4-Tiny algorithm for semantic object detection and then consolidated it with a 3D object pose estimation method and Kalman Filter to enhance the perception performance. In addition, a back-end UAV path planning for surveillance maneuver is integrated to complete the fully autonomous system. The perception module is assessed on a quadrotor UAV, while the whole system is validated through flight experiments. The experiment results verified the robustness, effectiveness, and reliability of the autonomous object tracking UAV system in performing surveillance tasks. The source code is released to the research community for future reference.

PDF Code

 

Learning-based Autonomous Inspection UAV system (LAIS)

The inspection of electrical and mechanical (E&M) devices using unmanned aerial vehicles (UAVs) has become an increasingly popular choice in the last decade due to their flexibility and mobility. UAVs have the potential to reduce human involvement in visual inspection tasks, which could increase efficiency and reduce risks. This paper presents a UAV system for autonomously performing E&M device inspection. The proposed system relies on learning-based detection for perception, multi-sensor fusion for localization, and path planning for fully autonomous inspection. The perception method utilizes semantic and spatial information generated by a 2-D object detector. The information is then fused with depth measurements for object state estimation. No prior knowledge about the location and category of the target device is needed. The system design is validated by flight experiments using a quadrotor platform. The result shows that the proposed UAV system enables the inspection mission autonomously and ensures a stable and collision-free flight.
 

PDF  Code