Research And Application Of Aerial Object Detection Algorithms Based On YOLOv3

The inspection of optical cable lines used to mainly be carried out by manpower,which is an inefficient way.The inspection of optical cable lines by UAV can greatly improve the efficiency and save human resources.Aiming at the requirement of automatic telecommunication cable line patrol by UAV proposed by a telecommunication company,this paper constructs an aerial object detection subsystem of UAV automatic line patrol system,which takes excavator as a risk object and detects the excavator in the returned photos of UAV to determine whether there is a risk in the photos.Through the analysis of 1004 5472×3648 high resolution aerial images provided by users and collected by themselves,it is found that the semantic information of such images is complex,the proportion of target objects is small,and the features are sparse,which makes the detection difficult and the detection easy to be disturbed.Users require that the recognition rate of excavator be more than 70% and real-time.According to this requirement,YOLOv3,which has the fastest detection speed and better detection performance,is selected as the object detection algorithm by comparing the current object detection algorithms.YOLOv3 object detection algorithm needs to input 416*416 image size,so when processing the input image,the image will be reduced,resulting in loss of features.In order to reduce the feature loss,a block-based YOLOv3 object detection algorithm is proposed.The algorithm uses a block method based on fixed step sliding window,divides the image into small blocks,and then uses YOLOv3 object detection algorithm to detect,which achieves the purpose of reducing the feature loss.Through comparative experiments,the recognition rate of block-based YOLOv3 object detection algorithm is nearly 10% higher than that without improvement.In order to train YOLOv3 object detection model better,a variety of aerial photographic excavator data sets are constructed,which lays a data foundation for future work.Based on these data sets,the critical value of digger indicators is explored.It is found that the recognition rate of digger is the highest when the pixel area of digger is 29143 and 18840.The YOLOv3 object detection algorithm based on block has been applied in actual production,and the aerial photograph object detection subsystem of the UAV telecommunication line inspection system has been developed.The subsystem has been deployed to the server and tested for more than four months,which has been recognized by users and has achieved the goal of improving work efficiency to a certain extent.