Research On Object Recognition Technology Of UAV Aerial Image Based On Deep Learning

Combining cameras with deep learning algorithms forms an intelligent processing system capable of performing object detection,and deploying it on embedded computing devices mounted on UAV allows for real-time analysis of aerial images captured by the UAV.This has significant market application value in the fields of intelligent transportation and security systems.However,in practical applications,aerial image analysis often requires high recognition accuracy and real-time performance.Additionally,to ensure the UAV’s endurance,there are strict requirements for size,weight,and power consumption,with numerous technical challenges to overcome.Firstly,the onboard computing platform needs to use embedded computing devices,but their low computing power often falls short;secondly,limited by the computational efficiency of ARM chips and NPUs used in embedded So Cs,the full precision inference time for models is long,which cannot meet real-time requirements;and the small size of objects in aerial images captured by UAV,along with their susceptibility to external environmental factors,further increases the difficulty of object detection.To enhance the speed and accuracy of aerial image object detection on resource-limited embedded computing devices,this thesis presents a deep learning-based UAV aerial image object recognition system built on NVIDIA Xavier.Utilizing SIMD technology to parallelize and accelerate image preprocessing algorithms,the system achieves more than 5times greater execution efficiency,addressing the significant time overhead associated with the implementation of preprocessing algorithms.By comparing the strengths and weaknesses of various algorithms in the context of embedded computing device application scenarios,the YOLOv3-tiny algorithm is chosen as the object detection algorithm deployed on embedded computing devices.TensorRT is used to accelerate model inference,and the system ultimately achieves over 90% accuracy with a single image inference time of 11.9ms,meeting real-time requirements.Experiments demonstrate that the system can perform the object detection tasks on UAV-mounted embedded computing devices as expected,with high accuracy,low resource usage,and maximum extension of UAV flight time.Moreover,the algorithm has strong portability,allowing it to be applied to the recognition and detection of other aerial targets based on the specific requirements of a given task.