Enhanced Attitude and Altitude Estimation for Indoor Autonomous UAVs

Vision-based particle filtering for quad-copter attitude estimation using multirate delayed measurements

In this paper, the problem of attitude estimation of a quad-copter system equipped with a multi-rate camera and gyroscope sensors is addressed through extension of a sampling importance re-sampling (SIR) particle filter (PF). Attitude measurement sensors, such as cameras, usually suffer from a slow sampling rate and processing time delay compared to inertial sensors, such as gyroscopes. A discretized attitude kinematics in Euler angles is employed where the gyroscope noisy measurements are considered the model input, leading to a stochastic uncertain system model. Then, a multi-rate delayed PF is proposed so that when no camera measurement is available, the sampling part is performed only. In this case, the delayed camera measurements are used for weight computation and re-sampling. Finally, the efficiency of the proposed method is demonstrated through both numerical simulation and experimental work on the DJI Tello quad-copter system. The images captured by the camera are processed using the ORB feature extraction method and the homography method in Python-OpenCV, which is used to calculate the rotation matrix from the Tello's image frames.

Functional Intelligence-Based Scene Recognition Scheme for MAV Environment-Adaptive Navigation

Adaptive navigation is the core of micro aerial vehicles (MAVs) conducting autonomous flights in diverse environments. Different navigation techniques are adopted according to the availability of navigation signals in the environment. MAVs must navigate using scene recognition technology to ensure the continuity and reliability of the flight. Therefore, our work investigated the scene recognition method for MAV environment-adaptive navigation. First, we exploited the functional intelligence-adaptive navigation (FIAN) scheme by imitating the physiological decision-making process. Then, based on sufficient environment-sensitive measurements from the environment perception subsystem in FIAN, the two-level scene recognition method (TSRM) in the decision-making subsystem consisting of two deep learning frameworks, SceneNet and Mobile Net-V2 was proposed to extract scene features for accurate diverse scenes recognition. Furthermore, the four-rotor MAV-Smartphone combined (MSC) platform simulating the owl’s omni-directional head-turning behavior was built. The proposed TSRM was evaluated for accuracy, delay, and robustness compared with PSO-SVM and GIST-SVM. The results of practical flight tests through MSC platform show that TSRM has higher classification accuracy than PSO-SVM and GIST-SVM, and performs smoothly with self-regulatory adaptations under diverse environments.