【作者】 李欢； 柴琳； 金立左；

【Author】 Huan Li;Lin Chai;Lizuo Jin;School of Automation,Southeast University;

【机构】 东南大学自动化学院；

【摘要】 本文主要对航拍车辆检测算法的改进问题展开研究,针对原始SSD对于大分辨率航拍图像中密集的小目标检测精度较低、实时性较差、资源占用较高等问题,本文提出一种基于改进的轻量级SSD的航拍车辆检测算法。Backbone由原来的VGG16替换为改进的MobileNetv2,可在保证检测精度的前提下大幅提升检测速度。通过对训练集的优化聚类,确定各预测层default boxes的尺度和纵横比,增强了对多尺度多纵横比的目标尤其是对小目标的检测能力。通过多尺度训练和大量的数据增强,大幅提升了泛化能力,并使模型的特征提取和检测能力得到充分发挥。在实验室航拍数据集上可同时保证检测精度和实时性,与SSD300相比,mAP提高了8.8%,与SSD512相比,mAP提高了4.5%,并且检测速度达到了25.1FPS,说明了本文算法对航拍视频中各类车辆目标检测的有效性和鲁棒性。更多还原

【Abstract】 Aiming at the problems that the original SSD has low accuracy, poor real-time performance, and high resource occupation for dense small targets in large-resolution aerial images, this paper proposes an aerial vehicle detection algorithm based on improved lightweight SSD. The original VGG16 is replaced by an improved lightweight network MobileNet v2 as backbone, which greatly improves the detection speed under the premise of ensuring detection accuracy. The scale and aspect ratio of the default boxes of each prediction layer are determined through the optimal clustering of the training set, which enhances the detection ability for multi-scale and multi-aspect-ratio targets, especially for small targets. Through multi-scale training and a large amount of data augmentation, the generalization ability is greatly improved, and the feature extraction and detection ability of the model are brought into full play. On the laboratory aerial data set, mAP is increased by 8.8% compared with SSD300, mAP is increased by 4.5% compared with SSD512, and the detection speed has reached 25.1 FPS, when the double guarantee of detection accuracy and real-time performance is achieved. The efficiency and robustness of the algorithm in this paper for the detection of various vehicle targets in aerial video are demonstrated.更多还原