基于改进YOLOV4-Tiny的UAV实时自动目标检测系统  

Real-time Autonomous UAV Object Detection System based on Improved YOLOv4-Tiny

作　　者：刘观生 谷峥 LIU Guan-sheng;GU Zheng(School of Information Engineering,Taizhou Vocational College of Science&Technology,Zhejiang,Taizhou 318020;Control Engineering College,Xinjiang Institute of Engineering,Xinjiang,Urumqi 310001)

机构地区：[1]台州科技职业学院信息工程学院,浙江台州318020 [2]新疆工程学院控制工程学院,新疆乌鲁木齐310001

出　　处：《贵阳学院学报(自然科学版)》2024年第4期83-87,94,共6页Journal of Guiyang University:Natural Sciences

基　　金：新疆维吾尔自治区自然科学基金项目“智能优化算法在农产品溯源系统中选址问题的研究”(项目编号:2020D01B20)。

摘　　要：为提高无人机(UAV)机载平台上目标检测的准确度和效率,提出了基于深度学习的UAV航拍图像实时检测系统。首先,使用轻量的改进YOLOV4-Tiny框架,利用移动倒置瓶颈模块优化骨干架构,提高存储效率,其次引入增强空间金字塔池化模块,在同一个卷积层中串联多尺度特征,改善网络的感受野。最后,向网络添加多尺度融合层以进行特征图合并,得到细粒度特征,从而提高对密集小目标的检测准确度。此外,利用TensorRT库对深度学习算法进行加速,进一步提高嵌入式平台上实时检测效率。实验中评估了UAV搭载不同单片机时的性能,并将所提检测算法与其他先进方法进行了比较。结果表明,所提系统实现了检测准确度和检测效率的最优平衡,在VisDrone和VEDAI数据集上的mAP性能比原始YOLOv4-Tiny分别提升了11.14%和8.9%。In order to improve the accuracy and efficiency of autonomous target detection on the unmanned aerial vehicle(UAV)platform,a real-time detection system for UAV aerial images based on deep learning was proposed.Firstly,a lightweight improved YOLOV4-Tiny framework is proposed,and the mobile inverted bottleneck modules are used to optimize the backbone architecture to improve storage efficiency.The enhanced spatial pyramid pooling module is introduced to connect multi-scale features in the same convolutional layer to enhance the receptive field of the network.Finally,a multi-scale fusion layer is added to the network to obtain fine-grained features,thereby improving the detection accuracy of densely distributed small objects.In addition,the TensorRT library is used to accelerate the deep learning algorithm,and the real-time detection efficiency on the embedded platform is further improved.In the experiment,the performance different single-board devices are evaluated,and the proposed detection algorithm is compared with other advanced methods.The results show that the proposed system achieves the optimal balance between detection accuracy and efficiency.The mAP performance of the proposed model on the VisDrone and VEDAI datasets is 11.14%and 8.9%higher than that of the original YOLOv4-Tiny framework,respectively.A detection speed of up to 42FPS is achieved with the proposed model on the single board device.

关 键 词：无人机 航拍图像 深度学习 移动倒置瓶颈 空间金字塔池化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]