Design Of UAV Vision-based Object Tracking System Based On Deep Learning And Binocular Vision

With the development of material technology,electronic technology,computer technology,the structure of UAV tends to miniaturized and intelligent,which makes UAVs have very wide applications in both military and civilian fields.Among them,the UAV object tracking is a very important category of UAV’s flight task,and is also an important research direction to achieve the fully autonomous UAV flight.UAV object tracking can achieve reconnaissance,tracking and other tasks in the military field;and can also assist in achieving dynamic aerial photography and other entertainment projects in the civilian field.In recent years,the UAV object tracking has become one of the hot issues in the research of UAV autonomous flight tasks.In this paper,we use a quadrotor UAV as the overall platform,a binocular camera as the main sensor,with deep learning,linear active disturbance rejection control and other schemes to achieve the design task of UAV object tracking.The design is divided into two parts,one is the specific three-dimensional spatial localization of specific target,and the other is the realization of UAV flight control and object tracking algorithm.For question 1,this paper uses a deep learning program to achieve dynamic target detection,and simultaneously perform target ranging by binocular vision,to obtain three-dimensional spatial localization of dynamic targets in camera coordinate system;for question 2,this paper designs a strong anti-interference and well-adapted flight control algorithm and object tracking algorithm.Finally,the effectiveness of this design scheme is verified by the outdoor flight test of the quadrotor UAV.The design scheme in this paper improves and solves some problems of target recognition accuracy,target localization speed and flight control algorithm robustness encountered in the UAV object tracking process.The deep learning target detection algorithm can identify objects with complex features,but there are some problems with computational complexity and recognition accuracy.This paper further improves the recognition accuracy by improving the deep learning module,while improving the network structure of SSD target detection algorithm reduces its computational complexity.Also,there is a large amount of invalid calculation in the operation of binocular vision algorithm,this paper obtains real-time target location information by extracting the running speed of the binocular vision algorithm by extracting the region of interest.Considering the characteristics of quadrotor UAV flight control model with multiple channels and strong coupling,this paper designs a flight control algorithm and object tracking algorithm based on linear active disturbance rejection control,which makes the quadrotor UAV have better anti-disturbance ability and robustness.