Real-time Detection Of Drone Video Object Based On Deep Learning

Drone-based video object detection has been widely applied in many fields such as military battlefield reconnaissance,urban emergency monitoring and intelligent traffic management.Object detection based on deep learning has become an important method and key technology for video object monitoring.However,the objects in drone images are usually small with drastical scale changes and low resolution,which makes high-precision object detection become very difficult.At the same time,multiple monitoring tasks with drones have high real-time requirements for video object detection.Aiming to solve the above problems,the real-time deep object detection algorithm for drone videos and the development of its software system are studied in this thesis,focusing on the limited sample data augmentation for network training,small object detection,parallel data processing and object detection system integration.The main research content and results are as follow:(1)Focusing on the degradation of detection performance caused by large scale changes of objects in drone images,a multi-scale sample augmentation method based on Poisson fusion is studied.The Poisson seamless fusion based on regional divergence consistency and boundary constraints is adopted to generate multi-scale samples with natural transition between the objects and the background.Experiments show that the proposed method effectively improves the multi-scale object detection performance,specifically,the detection rate is increased by 13.8% at 0.6 times the scale of the original image,and the detection rate is increased by 4.64% at 1.6 times the scale of the original image,which solves the limited sample problem during training and testing of deep neural network.(2)Focusing on the problem of low recognition precision of small objects,a cross-scale knowledge distillation model for small target detection is first constructed,which provides a novel learning mechanism for cross-scale knowledge transfer.And then the feature level alignment(FA),adaptive key distillation position algorithm(AKDP)and position-aware L2 loss are proposed to optimize the cross-scale knowledge distillation.By effectively utilizing knowledge distillation to enhance the feature extraction of small objects,experiments show that the proposed method has outstanding advantages of high detection accuracy of small objects compared to the other mainstream methods.Among them,the lightweight version only increases the calculation by 8.9%,significantly improving the average accuracy(21.62%)and small object detection accuracy(114.2%).(3)Using the parallel processing architecture design based on dictionary data structure,real-time(the processing speed is above 20FPS)video object detection processing was realized in this thesis;Developed a practical real-time drone video monitoring system,which has five functional modules including video analyzing,playing,detection,information and monitoring...