Study On UAV Autonomous Object Detection Technology Based On Airborne Vision

Multi-object detection technology based on computer vision,as an intelligent and convenient monitoring method,is more and more widely used in fields such as unmanned driving,urban safety,fire prevention and disaster relief.In the scene of drone shooting on the ground object,how to achieve high precision and real-time object detection is an important subject.Due to the particularity of the drone aerial photography scene,the object has a large scale in the image.Under different observation angles,the object is prone to morphological changes,scale changes,and illumination changes,resulting in poor robustness of detection algorithms.Therefore,the stability of the object detection algorithm faces a series of challenges.At present,the state-of-the-art object detection algorithms are based on deep learning,and are designed and debugged manually,requiring human experts to spend a lot of time to complete.At this stage,people are more and more interested in the process of implementing automated architecture design.To access the above problems,this paper uses the airborne vision UAV to collect the color images,and applies the deep neural network and neural network architecture search algorithm to carry out the study on the multi-object detection method of UAV on the ground,so as to realize the automatic design process of the architecture and achieve better performance.The object of this paper is the car in the complex environment.The specific research contents are as follows:1)This paper adopts an automatic architecture search algorithm,and continuously relaxes the search space,so that the search process reaches convergence faster.The connection between features is improved,and the attention module is introduced,which makes the model focus on more important features,reducing redundant information to a certain extent.2)This paper improves the pre-processing process of the input data,through filtering,adding noise,mosaic data augmentation and other methods to reduce the difficulty of subsequent recognition and classification and enhance the robustness of the model.3)This paper improves the alignment method of visual and semantic information,and introduces visual-semantic feature matrix.According to the method of human cognition,combining with the past experience,visual information and semantic dictionary to complete the object detection task,the detection accuracy can be improved while the object of the unseen classes can be better recognized.4)To verify the effectiveness of the multi-object detection method in this paper,a multi-object image recognition system is established.Using PyQt graphical development tools to encapsulate the algorithm of this article and design the system interface and functions,it is finally verified that the improved object detection algorithm performs better than the state-of-the-art object detection algorithms without experts to design the architecture.