基于MM YOLOv4的无人机目标检测算法

针对当前无人机目标图像检测方法精度较低和检测速度过慢的问题，提出一种结合轻量级网络和改进多尺度结构的目标检测算法。首先采用MobileNetV3轻量级网络替换YOLOv4的主干网络，减少模型复杂度，提升检测速度；其次，引入改进多尺度结构的PANet网络，增强高维图像特征和低维定位特征的流动叠加，提升对小目标的分类和定位精度；最后，利用K means方法对目标锚框进行参数优化，提升检测效率。同时结合公开数据集和自主拍摄方式构建一个新的无人机目标图像数据集Drone dataset，并基于数据增强的方法开展算法性能实验。实验结果表明，该算法的mAP达到了91.58%，FPS达到了55帧/s，参数量为44.39 M仅是YOLOv4算法的1/6，优于主流的SSD、YOLO系列算法和Faster R CNN算法，实现了对多尺度无人机目标的快速检测。

Research on UAV Target Detection Algorithm Based on MM YOLOv4

Aimed at the problems that the target image detection methods are low in accuracy and slow at detection speed of current UAV (Unmanned Aerial Vehicles, UAV), a target detection algorithm in combination with the lightweight network and the improved multi scale structure is proposed. Firstly, MobileNetV3 lightweight network is used to replace the backbone network of YOLOv4, reducing the model complexity and improving the detection speed. Secondly, the improved multi scale PANet network is introduced to enhance the flow superposition of high dimensional image features and low dimensional location features, and improve the classification and location accuracy of small targets. Finally, the K means method is introduced to optimize the parameters of the target anchor frame to improve the detection efficiency. Meanwhile, a new UAV target image dataset Drone dataset is constructed by combinating with the open dataset and the self shot images. The results show that the mAP and FPS of the proposed algorithm reach 91.58% and 55 f/s, and the parameter number of 44.39M is only 1/6 of the YOLOv4 algorithm and is superior to the mainstream SSD, the YOLO series and the Faster R CNN algorithms.