基于嵌入式平台与优化YOLOv3 的航拍目标检测方法

针对部署在嵌入式平台的目标检测模型在检测航拍目标时存在的检测速率低、耗时高、存储容量低的问 题，提出一种基于优化YOLOv3 算法的航拍目标检测方法。通过模型剪枝极大地减少了模型参数量，使用二分 K-means 对传统的锚框聚类算法进行优化改进，引入CIOU 损失函数加强边界框回归效果，再经TensorRT 对模型优 化加速后将该检测模型部署到JetsonTX2 平台上。选取大量不同类别不同环境的航拍图像制作数据集进行实验对比。 结果表明：优化后的算法在检验不同航拍图像目标时平均精度可达到83.9%，对每张图片的检测速度从2.8 FPS 提升 至14.7 FPS，满足精确性和实时性要求。

Aerial Target Detection Method Based on Embedded Platform and Optimized YOLOv3

Aiming at the problems of low detection rate, high time consumption and low storage capacity of target detection model deployed on embedded platform when detecting aerial targets, proposes an aerial target detection method based on optimized YOLOv3 algorithm. The number of model parameters is greatly reduced by model pruning. The traditional anchor box clustering algorithm is optimized and improved by using binary K-means. The CIOU loss function is introduced to enhance the effect of bounding box regression. After the model is optimized and accelerated by TensorRT, the detection model is deployed on JetsonTX2 platform. By selecting a large number of aerial images of different types and different environments to make data sets, the experimental results show that the average accuracy of the optimized algorithm can reach 83. 9% when detecting targets in different aerial images, and the detection speed of each image is improved from 2. 8 FPS to 14. 7 FPS, which meets the requirements of accuracy and real-time.