Research On Lightweight Of UAV Target Detection Algorithm Based On Deep Learning

With the rapid development of artificial intelligence technology,machine vision combined with deep learning has broader application prospects in the recognition and detection of industrial fields.In the task of target detection,with the continuous development of convolutional neural networks,the accuracy of target detection is getting higher and higher,and the cost is that the network model is getting bigger and bigger.Although the accuracy of target detection of convolutional neural networks exceeds that of traditional methods,it cannot meet the requirements of real-time and accuracy at the same time in small embedded devices such as UAV.How to improve the lightweight of convolutional neural network model has become a new research hotspot.Aiming at the target detection network based on deep learning for airborne embedded devices to ensure that it meets the requirements of real-time and accuracy at the same time,the main research contents of this thesis are as follows:First of all,this thesis discusses the development history of target detection technology of UAV and deep learning,introduces the relevant theoretical basis of deep learning,and analyzes the lightweight strategies of commonly used convolutional neural networks,which lays the foundation for subsequent research.Subsequently,this thesis deeply studies the SSD object detection network.The SSD network can predict targets of different scales on multiple feature scales,and has a good detection effect on small targets,which is suitable for UAVs to detect ground targets at high altitudes.After analyzing the SSD network structure,this thesis makes lightweight improvements based on the SSD network model.In order to balance the size and performance of the network model,the lightweight network MobileNetV2 is finally selected to replace the original VGG16 network of the SSD algorithm for deep feature learning.The deep separable convolution structure of mobilenetv2 combined with SSD network can effectively reduce the size of the network model,and then verify the effect of this algorithm on Pascal VOC data set.It is proved that the detection speed of this algorithm is significantly higher than that of the original SSD network model.Finally,this thesis analyzes the model of the quadrotor UAV,and studies the imaging principle of the camera and uses Zhang’s calibration method to calibrate the camera.Then combined with the robot operating system ROS,Gazebo physical simulation platform and PX4 on SITL,an unmanned aerial vehicle is built.The experiment realizes the application of the improved deep learning-based lightweight target detection algorithm in the UAV target detection task.