Research On Obstacle-avoidance Algorithm Of UAV Based On Improved Bisenet Model

With the continuous development and in-depth research of artificial intelligence and computer vision technology,the automation and intelligent flight control of UAV become possible.In the research of intelligent flight control of UAV,the ability to detect obstacles autonomously and avoid obstacles in real time has become the research hotspot of many scholars in recent years.Aiming at the obstacle avoidance mission of UAV,this paper studies the obstacle avoidance method of UAV based on deep learning without GPS and 3D map and only using monocular camera.The overall framework of the obstacle avoidance system designed in this paper consists of two parts:front-end environment perception and back-end mapping control.The front-end environment perception part uses the improved Bilateral Semantic Convolutional Neural Network(BiSeNet)to perceive the UAV flight environment information,and generate the possible collision probability between the UAV and the flight environment and the steering angle required for obstacle avoidance flight.The back-end mapping control part is composed of ROS system.Based on the first-order low-pass filtering algorithm,the forward flight speed and yaw angle mapping algorithm is designed,and the prediction information generated by the Bisenet model is mapped into the forward flight speed and yaw angle control instructions of the UAV.The main research contents of this topic are as follows:According to the task requirements of UAV obstacle avoidance,the overall scheme of UAV obstacle avoidance based on deep learning and computer vision is firstly designed.In order to make the obstacle avoidance system has universality and could be applied to a wide range of civil quadrotor UAV,while reducing the load of UAV and ensuring the endurance ability of UAV,this subject explores the obstacle avoidance system based on monocular camera.By studying UAV flight principle and deep learning theory,this paper designed an improved Bisenet model.The improved Bisenet model is designed by studying UAV flight principle and deep learning theory.The Bisenet model outputs the probability of collision between the UAV and the surrounding environment and the steering angle required by the corresponding obstacle-avoiding flight by inputting the real-time environment images taken by the on-board monocular camera of the UAV.Then,in order to generate real-time,sensitive and stable UAV control command,the working principle of low-pass filtering algorithm is studied.Based on the first-order low-pass filtering algorithm,the forward flight speed mapping algorithm and yaw angle mapping algorithm are designed,and the prediction information with spatial characteristics output by Bisenet model is combined with the time-domain information.The collision probability and steering angle generated by the Bisenet model were mapped into the forward flight speed and yaw angle control commands of the UAV respectively.This paper uses a variety of performance evaluation indexes,and in a variety of real environment to design the UAV obstacle avoidance scheme performance analysis and experimental verification.For the perception prediction of UAV flight environment images,the estimated value of collision probability generated by Bisenet model reached 99.2%accuracy and 0.968 F1 score.The root mean square error(RMSE)of the predicted value of the generated steering angle was 0.016,and the interpreted variance score was 0.837.The back-end mapping algorithm has high real-time performance,sensitivity and stability,which can quickly respond to the environmental perception information output by Bisenet model,filter and de-noise processing,and generate stable UAV control instructions.In this paper,the designed obstacle avoidance system is used to carry out many flight tests on the UAV in the real environment,which proves the feasibility and effectiveness of the designed obstacle avoidance system of the UAV.