For the challenge of limited computing resources of embedded UAV platform,in order to achieve high accuracy and lightweight real-time UAV aerial target detection,an improved YOLOv8 lightweight aerial target detection algorithm is proposed.The Separation Enhanced Attention Module(SEAM)is introduced to alleviate the problem of target occlusion in aerial images.A detection layer for small targets is added to improve the detection accuracy of small targets.The Ghost module is integrated into C2f module to form C2f-Ghost module,which significantly reduces the number of model parameters.The global channel pruning is performed on the improved network to ensure the detection accuracy and further reduce the model parameters.The pruned model is deployed on the Jetson Xavier NX embedded platform,and TensorRT is used to accelerate model inference.The experiment is performed on VisDrone2019 dataset and the comprehensive index is better than that of the comparison algorithm.The average detection accuracy reaches 53.6％,an increase of 4.1％.The number of model parameters and calculation amount are reduced by 88.4％and 50.1％respectively.The detection speed is 24.17 frame/s on the embedded platform,which verifies the effectiveness of the method.
万方数据
维普
