Research On Low Altitude Object Detection System Based On Deep Learning

With the improvement of people's living standards and the gradually opening of lowaltitude airspace,the audience of low-altitude aircraft is also increasing.How to ensure the safety of low-altitude airspace and achieve efficient supervision of low-altitude objects has become an urgent issue.This thesis designs a miniaturized low-altitude object detection system based on deep learning to help improve the ability to supervise the safety of low-altitude airspace.The system combines object detection and image semantic segmentation algorithms to detect low-altitude objects.When the detection object is an aircraft category and the size is larger than a certain threshold,the semantic segmentation algorithm is used to separate the object from the background.Aiming at the problem of low-altitude object detection,this thesis mainly performs the following work:(1)At first,for the low altitude object imaging pixel in photoelectric devices,less characteristics is not obvious characteristics,combined with public data sets and their own manual annotation constructs the size is about 10000 copies of object detection image data set,data sets to birds,kites,unmanned aerial vehicles(uavs),balloon,airplane of five kinds of common low altitude object as the goal.(2)Through experiments comparing YOLOv3 and YOLOv3 tiny algorithms,it is found that the former has high detection accuracy and slow detection speed,while the latter has low detection accuracy and fast detection speed.This thesis is based on the YOLOv3 object detection algorithm,combined with MobileNetv3,Dense and other network ideas,to ensure high detection accuracy,try to reduce the amount of parameters and calculations of the model.Train and test on the self-built image dataset using the improved network structure.The results show that compared with the detection of YOLOv3 and YOLOv3 tiny algorithms on small-size images,the average detection accuracy of the improved algorithm is 48.91%,which is significantly better than 37.13% of the YOLOv3 tiny algorithm,but the overall detection accuracy is still slightly lower than 63.7% of the YOLOv3 algorithm.Taking Jetson TX2 as the experimental platform,the average detection speed of the improved algorithm is 5.5 frames/s,which is between 12.3 frames/s in YOLOv3 tiny and 2.4 frames/s in YOLOv3,which meets the requirements of low-altitude object detection systems.(3)When the detection object is an aircraft category and the size is greater than a certain threshold,semantic segmentation is performed to separate the object foreground and background in the image.Based on the DeeplabV3+ algorithm,the network structure is optimized and the model is compressed.Experiments show that the improved algorithm greatly improves the detection speed without reducing the accuracy too much.(4)On the basis of improved low-altitude object detection algorithm and improved image semantic segmentation algorithm,a miniaturized low-altitude object detection system is designed and realized based on PyQt5 and Jetson TX2 hardware platform,which improves the detection and supervision level of low-altitude object.This thesis designs an object detection system based on deep learning,which can be deployed in miniaturized equipment,has high flexibility and maneuverability,and has good detection effect,which can effectively improve the supervision efficiency of lowaltitude objects and guarantee the safety of low-altitude airspace,and has high practical application value.