改进SSD无人机航拍小目标识别

Improvement of Small Target Recognition Algorithm of Aerial Photography Images Based on SSD

通过机载摄像头和嵌入式系统,结合计算机视觉技术将无人机发展为智能前端是一个重要的发展方向,而无人机航拍目标识别也逐渐成为研究人员的研究热点。但航拍图像具有背景复杂、目标较小、特征不明显等特点,且受到嵌入式平台内存和计算能力的限制,在无人机上对航拍目标自动检测非常困难。论文提出了一种改进SSD算法,用宽残差网络(WRN)代替了原来的VGG16特征提取网络,降低了模型的参数量;针对训练样本中正负样本不匹配的问题,采用焦点损失函数,使模型训练中更侧重于检测困难样本。实验结果表明,该模型在VisDrone2018-Det数据集上达到0.76mAP,同时在Jetson TX2嵌入式平台上达到16FPS,准确率和实时性上均有了明显提高。

Through the airborne camera and embedded system,combining computer vision technology to develop UAV into in⁃telligent front-end equipment is an important development direction,and UAV aerial photography target recognition has gradually become a hot spot to researchers.However,aerial image has the characteristics of complex background,small target and inconspicu⁃ous features,and limited by the memory and computing ability of embedded platform,it is very difficult to detect aerial target auto⁃matically on UAV.In this paper,an improved SSD algorithm is proosed,using wide residual network(WRN)instead of the original VGG16 feature extraction network to reduce the parameters of the model.Focus loss function is used to solve the problem of mis⁃match between positive and negative samples in training process,which makes the model training more focused on the detection of difficult samples.The experimental result shows that the model achieves 0.76mAP on VisDrone 2018-Det dataset and 16FPS on Jet⁃son TX2 embedded platform.The accuracy and real-time performance of the model are improved significantly.