Research On Object Detection Of Small Drone Inspired By Biological Vision

UAV detection in complex background is always a difficult problem in the field of object detection,The classical object detection algorithms have been replaced by deep convolutional neural networks due to the disadvantages of large computation and low detection accuracy.The deep learning algorithms based on neural networks make use of powerful computing power and big data to realize the detection task of the UAV in simple background.However,when the UAV appears in the mountains,jungles,buildings or occluded in other complex backgrounds,the general detection algorithms of UAVs are difficult to make use of the target motion information,resulting in missed or false detection.To solve such problems,this paper uses infrared optical images with high stability and strong antiinterference to conduct drone detection.Building on rational and reliable analysis of the above existing problems,the main contributions of the thesis are as follows:Firstly,there are few publicly available UAV datasets.Drones with different shapes and sizes have different imaging characteristics in infrared images.As for deep learning algorithms,different feature information will cause great differences in the accuracy of algorithm detection.In this paper,the infrared video drone dataset Anti-UAV2020 is augmented by digital image processing methods such as rotation,scaling and adding noise,to ease the problem of sample shortage.At the same time,the improved YOLO-v3 algorithm is verified and tested on this augmented dataset.Secondly,because of the defects of convolutional neural networks,deep learning algorithms have poor detection ability for small drones in complex backgrounds.Due to the small proportion of pixels of small drone in infrared video and the low contrast between drone and background in complex background,the texture information of drone in the image cannot be extracted effectively,which leads to the failure of detection algorithm.Inspired by the visual information processing mechanism of primates,this thesis utilizes the advantages of biological vision to observe moving objects in a complex background,and then mathematically models the signal transduction process of retinal cells.The visual signal processing process of retina is partly simulated by algorithms.The brain-like computing model is used to reduce the noise of the UAV in the complex background.The magnocellular pathway is introduced to extract the motion information of the UAV,and in this way,the features of the UAV are enhanced.Subsequently,a feature fusion step is carried out with the classic real-time target detection algorithm YOLO-v3,We propose a new UAV video target detection algorithm M-YOLO(Magno-YOLO).The performance of the proposed model is tested and evaluated based on the publicly available dataset.Thirdly,the drone video detection algorithm is embedded to a system consisting of an infrared camera,an tripod head and a high-performance processor hardware devices.Therefore,a prototype that can test and demonstrate the designed drone detection algorithm is built.The performance of the proposed UAV detection model is further tested under real-world complex backgrounds.