Research And System Implementation Of Dynamic Small Target Detection And Tracking Technology Based On Deep Learning

With the development of the drone industry and a higher demand for an intelligent society,drones are highly used in different scenarios in these days,including aerial photography,agricultural industrialization,public security fire supervision and logistics and warehousing.However,some other drone users still cannot use the drone properly and their lack of safety and legal awareness leads the inappropriate and even illegal activities of drones happen from time to time.The corresponding supervision technology is in need urgently.The existing algorithms towards drones,including visual detection and target tracking,are not yet perfect.Most dynamic small target detection methods for drones rely too much on the detection capabilities of the traditional motion information detection,limiting the drone detection and tracking capabilities in a dynamic background.In order to solve this problem,this paper conducts research on the detection and tracking technology of low altitude small drone targets,designs and implements a visual perceptionbased drone photoelectric detection and tracking system.The convolutional neural network is used in the algorithm.The real-time detection and tracking of the drone is realized and the dependence of the detection model on traditional motion information detection methods is reduced.This paper is mainly divided into three parts and the specific contents are as follows:1.First,the moving target detection method for dynamic small targets of drones is discussed.Also,the classic tracking algorithm for dynamic small targets of drones is introduced,The theory of the algorithms and the experiment results are discussed.I summarize the advantages and disadvantages of the algorithm and the feasibility analysis used in the design.2.To solve the problem of large-scale changes,complex backgrounds difficult to detect and real-time detection in the traditional algorithms when facing drone problem,this paper designs a real-time detection algorithm based on neural network model.First,use the classic moving target detection method to scan the suspicious areas,and then use the deep learning-based real-time target detection algorithm to perform dynamic small target detection task of drones.The deep learning real-time detection algorithm is performed in two steps.First,neural networks are used locally,performing super-resolution reconstruction of image content to enhance detailed information.Then use the improved YOLOv3 neural network model for drone detection.The experimental results show that the method in this paper has good real-time performance under the acceleration of the graphics processor,has high accuracy and recall rate of the detection,good robust.Compared with traditional algorithms and machine learning type feature extraction algorithms,the processing speed is faster and the detection is more stable.3.For the possible problems of classic tracking algorithm for dynamic small targets of drones,this paper designs the corresponding tracking algorithm process.When the number of targets is small,Kalman filter is used to predict the position of the next frame of the target,super-resolution reconstruction method is used to strengthen the local information,then the kernel-related filter tracking algorithm KCF is used to track the detected target.When the number of targets is big,the detection tracking is used to track the target in real time.First,use the detection module to perform full-image target detection,and then perform target detection in a new frame.Use simple online real-time tracking algorithm to associate the target tracking frame between two frames,determines the relationship of the borders in the previous and subsequent frames,and realizes multi-target tracking.The experimental results show that the algorithm can effectively track dynamic small targets and performs well and meets single-target and multi-target tracking requirements with real-time performance.4.Design and implement the drone detection and tracking system,implement the previously described process in reality scene,and design a software to achieve the control of image and video extraction equipment,target detection and tracking functions.Introduce the overall structure of the system again,and modularize the components involved in the system.Implement and integrate each module into a system.Finally verify the effectiveness of the system through actual scenario test.The effective distance can reach 450 meters with high accuracy and good real-time performance.