| Multi-rotor UAVs are widely used for reconnaissance,surveillance,targeting and battlefield damage assessment due to their low cost,operational flexibility and high level of intelligence.Accurate identification of multi-rotor UAVs is crucial to grasp the battlefield initiative,improve the performance of future command and control systems and strike targets with precision.In this paper,we take multi-rotor UAVs as the research object and investigate the multi-rotor UAV detection algorithm based on Yolov5.The main research elements of the paper are as follows:1)To address the current situation that existing multi-rotor UAV datasets with high sample quality are scarce and difficult to obtain,this paper integrates the datasets obtained from the Kaggle platform and optimizes the multi-rotor UAV datasets by data augmentation to form a new multi-rotor UAV dataset.2)An improved Yolov5 multi-rotor UAV detection algorithm is investigated to improve the detection effectiveness of existing multi-rotor UAV detection algorithms.First,a co-attentive mechanism is introduced in the Yolov5 backbone part,thus improving the feature representation of the target of interest in the image and suppressing the interference of background noise information.Then,adaptively spatial feature fusion techniques are injected into the head of the baseline model to facilitate the fusion of feature maps with different spatial resolutions in the model.Finally,the original bounding box regression loss function in Yolov5 is optimized by replacing the original loss function with the SIo U loss function,and by introducing the covariate of angular loss into the metric of the previous loss function,the degrees of freedom of the prediction box to move during the training process are reduced to improve the accuracy of the prediction box repositioning.3)Ablation experiments based on the new multi-rotor UAV dataset were conducted using the base model Yolov5 and the improved model respectively,and an attempt was made to introduce more similar techniques in the current computer vision field as a comparison.The experimental results show that the improved Yolov5 algorithm can achieve an average accuracy of 97.37%,which is 4.59% higher than the original model Yolov5,and the recall rate can reach 95.23% with an accuracy of 96.9%,which is 1.94% higher than the initial model,and the detection effect has been significantly improved. |
