Research On Binocular Vision Technology In Obstacle Recognition Of Quadrotor UAV

UAVs have the advantages of small size,light weight,simple mechanical structure,low cost and strong maneuverability,and have high military and civilian values.UAVs need a variety of algorithms and systems to assist them during navigation.According to the task content,the drone software system can be divided into four major modules: perception,cognition,decision-making and control.Among them,the sensing module is mainly used to obtain environmental information,which is the basis for the UAV to achieve autonomous or semi-autonomous flight.Obstacle recognition technology is the most important and basic part of perception.UAV obstacle recognition technology is a technology in which the drone can autonomously identify obstacles and determine the distance between the obstacle and its own body during the flight.This technology can greatly reduce crashes during the drone flight.Obstacle recognition technology based on computer vision has gradually become a research hotspot because of its hardware price advantage,and has high research and application value.In this paper,the objective of UAV obstacle detection based on binocular vision and target detection is studied.The visual measurement method mainly based on the binocular method is studied,and the functions of calibration,correction and preprocessing of the binocular system are realized based on Open CV..In binocular vision,this paper studies and compares various parallax algorithms.In view of the existing methods that rely on fixed windows and have a poor effect on low-texture areas,this paper proposes an adaptive window using gradient features of Census.The parallax algorithm improves the effect of stereo matching.In the aspect of target detection,a deep learning-based target detection method is studied.In view of the existing methods,the problems of insensitivity to small targets,high time-consuming,complex network structure,and large computational load Based on the limited computing power of the UAV's airborne embedded mobile platform,this paper improves the existing deep learning-based target detection methods.By increasing the scale of the YOLO v3-tiny model,using feature fusion modules,and adjusting the number of layers,etc.Adjusted and proposed the YOLO v3-tiny-stronger target detection model.In addition,this article is based on a four-rotor drone of a logistics company,using NVIDIA Jetson TX2 as an airborne computing platform,with DJI N3 flight control system and MYNT S2110-95 / Color type binocular camera module to build a four-rotor Experimental platform of human-machine obstacle recognition system,and realizes on-board deployment of required equipment and algorithmsBased on this,this paper combines the proposed parallax algorithm and target detection algorithm to design a UAV obstacle recognition algorithm flow based on binocular vision and target detection,and designs related experiments to verify the algorithm in this paper.The experimental results show that in low-altitude scenarios,the drone obstacle recognition algorithm designed in this paper can detect obstacles on the flight path of the drone and calculate the distance between the obstacle and the drone in real time.Because the actual distance of obstacles in front cannot be accurately tested during the drone flight,this article experimentally verifies the accuracy of the algorithm's distance measurement of obstacles under ground conditions.The experimental results show that for obstacles within 10 meters The detection error of this system is less than 5%,which basically meets the accuracy requirements of obstacle recognition for drones.